Daily 2026 · 06 · 25
Tech Economic

Qualcomm puts a CUDA-agnostic compiler stack under its data-center play, paying $3.92 billion in stock for Modular and naming Meta as its first big CPU customer

Qualcomm closed Investor Day on June 24 with an all-stock $3.92 billion acquisition of Chris Lattner's Modular at 19.2 million shares, a new Dragonfly C1000 250-plus-core 5-gigahertz server CPU with a multi-year, multi-generation Meta Platforms supply agreement, and a Dragonfly AI300 accelerator built for liquid and air cooling, alongside fiscal-2029 non-handset revenue guidance lifted to $40 billion from $22 billion and adjusted EPS to $18 against a $15.26 consensus. The stock jumped 14 percent in after-hours trading and reports surfaced of separate Qualcomm talks to buy Tenstorrent at $8 to $10 billion.

SiliconANGLE · CNBC · Modular · 2026-06-24 · Source ↗ ·

Qualcomm closed its annual Investor Day on Wednesday, June 24 with three announcements that reorganise the company's pitch from a smartphone-modem business with a side data-center bet into a vertically integrated agentic-AI platform. The first is the all-stock $3.92 billion acquisition of Modular Inc., the AI infrastructure software company founded by LLVM and Swift creator Chris Lattner. Qualcomm is paying 19.2 million of its own shares for the company, a price that more than doubles Modular's most recent funding round of $380 million from GV and Greylock. The deal is expected to close in the second half of 2026, subject to customary regulatory approvals. The second is the formal launch of a new server-CPU family called Dragonfly C1000, a 250-plus-core processor running above 5 gigahertz with what the company described as roughly double the performance-per-watt of comparable server processors. Qualcomm named Meta Platforms as the multi-year, multi-generation customer who will deploy the C1000 in production, without disclosing the financial terms. The third is a refresh of the company's data-center accelerator roadmap under the Dragonfly AI300 name, a machine-learning chip designed to slot into either liquid-cooled or air-cooled racks. The combined package is what gave the stock a 14 percent jump in after-hours trading and what let CFO Akash Palkhiwala lift fiscal-2029 guidance for non-handset revenue to $40 billion from a previously communicated $22 billion, with adjusted earnings per share at $18 against a consensus of $15.26.

Why Modular matters more than the price tag suggests

Modular is best understood as a compiler-stack bet against Nvidia's CUDA moat. The company's MAX platform is a vendor-neutral runtime that lets developers write inference code once and execute it across heterogeneous hardware, Nvidia GPUs and AMD chips and Intel processors and Qualcomm NPUs and custom ASICs, without rewriting per target. The companion language Mojo extends Python with the kind of low-level control that Triton or CUDA-C provide on Nvidia hardware. The strategic logic is that any non-Nvidia compute platform faces the same headwind: customers who have written their model-serving code against CUDA's primitives cannot move that code to a different stack without re-tuning. Modular's pitch is that the porting layer is a product, and that the platform supplier who controls it owns the migration path away from Nvidia. Buying it inside Qualcomm gives the Dragonfly C1000 and AI300 a software story they did not have last quarter, and it lets Qualcomm sell the Modular layer to customers who buy other accelerators while still cashing the Qualcomm-margin software dollar.

"As agentic AI scales across data centers and edge environments, the industry is moving toward disaggregated, multi-vendor architectures that demand a more open and modern software foundation."— Cristiano Amon, President and CEO, Qualcomm

Lattner reframed the move in personal terms in the Modular blog post that went up immediately after the announcement. He has been here before: LLVM at Apple, then Swift, then a stint inside Google running Brain's hardware-software interface team, then a year on the Tesla autopilot team. The thesis at each stop has been the same. Compiler infrastructure is where platform power gets cashed. Owning the runtime that everybody else's models route through compounds.

"Joining Qualcomm gives us the scale and platform reach to accelerate that mission. To make AI development more accessible and performant."— Chris Lattner, Co-founder and CEO, Modular

Dragonfly C1000, the Meta supply deal, and what the fiscal-2029 number means

The Dragonfly C1000 specification is aggressive. Two hundred and fifty cores at sustained clocks above five gigahertz puts the chip in the high-end x86 server territory currently held by AMD's Bergamo and Sienna SKUs and Intel's Sierra Forest, with a performance-per-watt advantage that Qualcomm pegged at roughly two times comparable parts. The chip is Arm-based, derived from the Nuvia design team Qualcomm acquired in 2021. Meta is the first named customer. The supply agreement is multi-year and multi-generation, which is the structure that ARM-based server CPUs have been seeking since the early Graviton days at AWS: a hyperscaler committing to a roadmap, not a one-off purchase. Financial terms were not disclosed. The Dragonfly AI300 sits on the accelerator side. It is the same chip family Qualcomm previewed earlier this year as AI200 and AI250, now renamed and re-positioned alongside the CPU under a single Dragonfly brand. The accelerator is engineered for both liquid and air-cooled deployment, which matters for the second-tier customers who cannot economically rebuild their existing data halls for direct-liquid only.

The fiscal-2029 guidance change is the part that priced into the stock. Non-handset revenue jumped to a $40 billion target from $22 billion previously, almost doubling the trajectory inside the same window. Adjusted EPS at $18 sits roughly 18 percent above the analyst consensus of $15.26. Qualcomm did not specify how the $40 billion splits between automotive, IoT and data center, but the bulk of the new headroom is data center: there is no other business line growing fast enough to add $18 billion of guidance in three fiscal years.

The Tenstorrent rumour and the broader pattern

Reuters and Bloomberg reported during the day that Qualcomm is in separate talks to acquire AI chip startup Tenstorrent for $8 to $10 billion, several times the price Qualcomm just paid for Modular. Tenstorrent is the RISC-V-based AI compute company led by Jim Keller, and it has been the most credible non-Nvidia hardware bet outside the hyperscalers' internal ASIC programs. The pattern across the day reads as Qualcomm spending what it needs to spend, in equity, to assemble the full vertical stack: CPU and accelerator silicon for the hardware layer, Modular for the runtime, RISC-V optionality at the architecture layer, Meta as the anchor customer at the demand layer. Whether the Tenstorrent deal closes is the open question that determines whether the architectural breadth claim holds.

The competitive read for the rest of the industry is mixed. Nvidia keeps the CUDA installed base and the training market. AMD and Intel keep their x86 footprint. The pressure lands on the second-tier accelerator startups, who have largely competed on price and on per-workload performance and who now face a Qualcomm with shipping hyperscaler customers and a compiler runtime under one roof. The pressure also lands on the hyperscaler internal silicon programs at Google, AWS, Microsoft and Meta, all of whom face the same compiler problem Modular was built to solve. Meta's decision to take the C1000 first is the most informative single fact in the announcement: it suggests Meta sees value in a merchant-silicon partner with a credible runtime, not only in extending its own MTIA and Athena programs.

Big Picture

Qualcomm spent the day reorganising its pitch from a smartphone-modem business into a vertically integrated agentic-AI platform, paying $3.92 billion in 19.2 million of its own shares for Chris Lattner's Modular to put a CUDA-agnostic compiler stack under a new Dragonfly C1000 250-core 5-gigahertz server CPU with Meta as the multi-year multi-generation anchor customer, refreshing the Dragonfly AI300 accelerator for liquid or air cooling, lifting fiscal-2029 non-handset revenue guidance to $40 billion from $22 billion and adjusted EPS to $18 against a $15.26 consensus, sending the stock up 14 percent after hours, and reportedly negotiating a parallel $8 to $10 billion acquisition of RISC-V-based Tenstorrent. Micron printed its own validation of the AI memory cycle three hours later with a $41.45 billion fiscal Q3 that quadrupled revenue year-on-year on $28.2 billion of profit, lifted Q4 guidance to a $49 to $51 billion range, and announced a strategic supply agreement covering HBM, DRAM and SSDs to Anthropic that includes Micron's participation in Anthropic's Series H round.

OpenAI broadened the Trusted Access for Cyber tier on June 24, naming Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, Oracle, Palo Alto Networks and Zscaler as commercial partners and Australia, Canada, France, Germany, Japan, South Korea and EU institutions as government partners, with GPT-5.5-Cyber scoring 85.6 percent on CyberGym against 81.8 percent for standard GPT-5.5, 39.5 percent on ExploitGym against 25.95 percent, and 69.8 percent on SEC-bench Pro. Patch the Planet, the companion open-source-vulnerability programme run with Trail of Bits and HackerOne, has now publicly enrolled cURL, NATS Server, pyca/cryptography, Sigstore, aiohttp, the Go project, freenginx, Python and python.org and reports hundreds of issues identified plus dozens of merged patches across the first five-day sprint.

Hugging Face Daily Papers crowned Qwen-AgentWorld at 86 upvotes, the first language world model that simulates seven agent environments through long chain-of-thought reasoning, with the 397B-A17B variant scoring 58.71 on AgentWorldBench against GPT-5.4 at 58.25. Frontis AI's NatureBench at 48 showed the strongest coding agent surpasses Nature-family SOTA on only 17.8 percent of 90 cross-discipline tasks. Kwai filed two top-rated mobile-agent papers, MobileForge at 34 on annotation-free GUI adaptation and MemGUI-Agent at 33 on cross-session memory. OpenThoughts-Agent at 27 from a Stanford-Berkeley-NYU-CMU-UW-UCSD-ASU-UCLA-UNC-TUM-LAION collaboration shipped 15,000 SFT traces and 720 RL tasks across NL2Bash and InferredBugs to post-train Qwen3-8B for Terminal-Bench and SWE-Bench. Tsinghua AIR's AOHP at 25 landed an open-source OS-level agent harness. Shanghai AI Lab's EventVLA at 17 put event-driven visual memory into long-horizon vision-language-action policies. SAILING Lab at CMU and MBZUAI published Critique of Agent Model at 10.

MWC Shanghai opened the same day with humanoid robots walking the stage before any human speaker. GSMA Director General Vivek Badrinath declared the mobile industry's mission formally shifted from connecting people to enabling intelligent machines. China Telecom signed a Robotics-as-a-Service deal with AgiBot, the Shanghai company that crossed 10,000 humanoid robots shipped in March and led 2025 humanoid shipments at 39 percent global share per Omdia. The Mobile AI Innovation Frontiers Zone debuted a Humanoid Robot Football Penalties Challenge that bars pre-programming, hidden controllers and resets. The UK Financial Conduct Authority's Nikhil Rathi told Agents of Change that legislation will never keep up with AI, naming stewardship as the FCA's new operating posture and announcing agentic-AI exploration as a wholesale-market first responder, a Mills Review on AI in retail finance, an AI practice-guidance publication later in 2026 and a Supercharged Sandbox built with Nvidia and Google. Microsoft's third annual AI in Education Report put 92 percent of students and education leaders on AI for school against a methodology of 3,345 respondents across six countries, alongside Copilot Notebooks, Study and Learn Agent and Unit Plans in Teach shipping at no extra cost on Microsoft 365 Education A3 and A5. The DOE's National Nuclear Security Administration unveiled Aires Tide, the first AI-designed flight test vehicle under the Genesis Mission. Polymarket traders abandoned the GPT-5.6 launch window for June 22 to 28 as odds collapsed from 83 percent earlier in the month to 18 percent, with the July window now pricing at roughly 94 percent. Uber capped employee AI-tools spend at $1,500 a month after exhausting its 2026 AI budget in four months. llama.cpp tagged build b9782 the morning of the 24th.

More today

17 stories · 2026-06-24
EconomicTech

Micron quadruples revenue to $41.45 billion as it locks a strategic HBM, DRAM and SSD supply pact with Anthropic and invests in the Series H

TechCrunch · Yahoo Finance · 2026-06-24 · Source ↗

Fiscal Q3 revenue of $41.45 billion quadrupled year-on-year, profit hit $28.2 billion against $1.88 billion a year ago, the stock crossed $1,200 with a $1.2 trillion market cap, and Q4 guidance set a $49 to $51 billion range. Micron also announced a multi-year HBM, DRAM and SSD supply agreement to Anthropic with a strategic investment in the Series H.

Micron Technology reported fiscal third-quarter results after the market close on Wednesday, June 24, posting revenue of $41.45 billion and profit of $28.2 billion. Both numbers quadrupled the year-ago period. Revenue a year earlier was roughly $10 billion, profit roughly $1.88 billion. The stock rose more than 13 percent in after-hours trading and traded above $1,200 a share into Thursday's open, taking the company's market capitalisation past the $1.2 trillion mark. Micron guided Q4 to a range of $49 to $51 billion of revenue, a sequential lift in the high single digits to mid teens off an already record quarter. The driver across the report is high-bandwidth memory for AI accelerators, where Micron sells into Nvidia's H200 and Blackwell platforms, AMD's MI300 family, and the internal accelerator programs at the hyperscalers. CFO Mark Murphy described the supply environment as fully sold out into 2027 with pricing power on every contract renewal.

The accompanying corporate news is the Anthropic agreement, made public on Tuesday, June 22 and tied into the Wednesday earnings narrative. The two companies announced a multi-year strategic supply agreement covering high-bandwidth memory, DRAM and solid-state drives for Anthropic's AI data-center buildout, with Micron also taking a strategic stake in Anthropic's Series H round at the $965 billion post-money valuation. Neither company disclosed the dollar amount of the Micron investment or the volumes in the supply agreement. The structural read is that hyperscaler and frontier-lab capital is starting to land directly on the memory-chip side of the supply chain, paying not only for guaranteed access but for a seat at the table on multi-generation roadmaps. Anthropic in particular needs HBM4 and HBM4E generations starting in 2026 and 2027 to populate the post-Blackwell systems it is leasing from SpaceX's Colossus 2 and from AWS, and a co-design relationship with the largest US memory supplier gives Anthropic a hedge against the Samsung and SK Hynix HBM concentration that has been the supply bottleneck of the past four quarters.

"Demand for high-bandwidth memory is fully booked into 2027, and we are co-designing the next two generations with our largest customers."— Mark Murphy, CFO, Micron Technology

The trading context is the South Korean memory complex that took $3.8 billion of foreign net selling and circuit-broker triggers on Tuesday on rumours of a US export-control package targeting HBM. Micron, as the only large US-headquartered HBM supplier, is the immediate hedge if any version of those rules materialises. Apple CEO Tim Cook warned consumers on the same earnings call cycle of unavoidable consumer-product price increases on the back of memory constraints, a signal that the bottleneck is now meaningfully past the data center and into general-purpose memory.

TechSafety

OpenAI graduates GPT-5.5-Cyber to broader Trusted Access and ships Patch the Planet into nine flagship open-source projects

Dataconomy · OpenAI · 2026-06-24 · Source ↗

GPT-5.5-Cyber scores 85.6 percent on CyberGym against 81.8 percent for standard GPT-5.5, 39.5 percent on ExploitGym against 25.95, and 69.8 percent on SEC-bench Pro, gated through Trusted Access for Cyber to Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, Oracle, Palo Alto Networks, Zscaler and government partners across seven jurisdictions.

OpenAI moved its specialised cybersecurity model GPT-5.5-Cyber out of limited preview on Wednesday, June 24, broadening the set of organisations that can apply for Trusted Access for Cyber and publishing the underlying benchmark numbers. On CyberGym, the agent-vulnerability discovery benchmark, GPT-5.5-Cyber scores 85.6 percent against 81.8 percent for standard GPT-5.5. On ExploitGym, which measures end-to-end exploit construction, the model scores 39.5 percent against 25.95 percent. On SEC-bench Pro, which evaluates security-engineering tasks, the score is 69.8 percent. The commercial partners cleared into the Trusted Access tier as of Wednesday are Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, Oracle, Palo Alto Networks and Zscaler. The government partners are Australia, Canada, France, Germany, Japan, South Korea and EU institutions. OpenAI continues to recommend standard GPT-5.5 as the default cybersecurity model for most workflows, with GPT-5.5-Cyber reserved for the more permissive end of the Trusted Access programme.

Alongside the model graduation, OpenAI shipped Patch the Planet, the Daybreak-branded open-source security programme run jointly with Trail of Bits and HackerOne. Patch the Planet pairs full-time security engineers from Trail of Bits with AI-assisted vulnerability research, then routes findings through HackerOne for coordinated disclosure. The initial set of projects publicly enrolled is cURL, NATS Server, pyca/cryptography, Sigstore, aiohttp, the Go project, freenginx, Python and python.org, with more than 30 projects committed to participate. The five-day initial sprint reported hundreds of security issues identified and dozens of patches merged, with many more findings still going through coordinated disclosure. Participating projects receive access to ChatGPT Pro, conditional access to Codex Security, and API credits earmarked for core development, maintainer automation and release workflows.

"GPT-5.5 remains our primary recommended model for most cybersecurity workflows through Trusted Access for Cyber. GPT-5.5-Cyber is designed to provide more permissive access for a smaller set of highly verified defenders and organizations."— OpenAI Trusted Access for Cyber announcement, June 24

The architectural design of Trusted Access is the part worth attending to. OpenAI's standard policy stack restricts capabilities like vulnerability identification, malware analysis and binary reverse engineering with classifier-based refusals, because the same capabilities that defenders use overlap with the capabilities that attackers use. Trusted Access flips that envelope for vetted organisations: vulnerability identification, triage, malware analysis, binary reverse engineering, detection engineering and patch validation pass through with reduced refusal rate, while requests that look like credential theft, stealth, persistence, payload deployment or unauthorised exploitation of third-party systems continue to refuse. The vetting layer is the gate. Patch the Planet is the productisation of the policy: a defined defender set, a defined open-source-project surface, an external security partner that owns the human-review step, and a disclosure pipeline that moves the output back into upstream packages.

The competitive context is Anthropic's parallel cybersecurity program, which has been offline since June 12 along with the rest of Fable 5 and Mythos 5 under the export-control directive, and Google's Project Zero plus Big Sleep automated-vulnerability programme inside DeepMind. OpenAI is moving on the empty space.

Tech

Qwen-AgentWorld posts a 35B and a 397B-A17B language world model that simulates seven agent environments and edges past GPT-5.4 on AgentWorldBench

arXiv 2606.24597 · Hugging Face Daily Papers · 2026-06-24 · Source ↗

The 397B-A17B model scores 58.71 on AgentWorldBench against GPT-5.4 at 58.25, simulating MCP, Search, Terminal, Software Engineering, Android, Web and OS environments through long chain-of-thought, trained in three stages of CPT, SFT and rubric-and-rule RL.

The Qwen Team's Qwen-AgentWorld paper at 86 upvotes is the first language world model to simulate seven distinct agent environments through long chain-of-thought reasoning. The seven covered are MCP, Search, Terminal, Software Engineering, Android, Web and OS. The release ships two model sizes: Qwen-AgentWorld-35B-A3B and Qwen-AgentWorld-397B-A17B, both mixture-of-experts. The training recipe is three-stage. CPT, continued pre-training, injects state-transition dynamics and world knowledge from environment interaction logs. SFT activates next-state-prediction thinking patterns. RL with hybrid rubric-and-rule rewards sharpens simulation fidelity against ground-truth environment behaviour. The companion benchmark AgentWorldBench is built from real environment interactions of frontier proprietary models such as Claude Opus 4.6 on widely used agent benchmarks including Terminal-Bench 1.0 and 2.0 and OSWorld-Verified, an explicit out-of-distribution evaluation set.

The headline number is 58.71. Qwen-AgentWorld-397B-A17B beats every frontier proprietary model the authors evaluated on the unified AgentWorldBench score, edging GPT-5.4 at 58.25 and beating Claude Opus 4.6 and Gemini 3.5 Pro variants. The 35B-A3B variant, with only 3 billion active parameters, sits below the frontier models on the overall score but inside the same general envelope on individual environments where state-transition complexity is lower. The Qwen team frames the work as the first time a small or mid-size open model has fully simulated an agent environment instead of just producing the next action inside one.

The downstream consequence is the part that matters for the broader research direction. World models are the substrate for offline RL and for planning at training time. If a language world model can simulate a terminal session, a browser or an Android UI faithfully enough that a separate policy model can train against the simulator instead of against the live environment, the cost of training capable agents drops by the cost of all the environment interactions. The paper hints at exactly that direction in its experiment design. Both models, the benchmark and the AgentWorldBench evaluation set are available on Hugging Face and GitHub at the QwenLM/Qwen-AgentWorld and Qwen/AgentWorldBench repositories.

Tech

NatureBench shows the strongest coding agent matches Nature-family SOTA on only 17.8 percent of 90 cross-discipline ML tasks

arXiv 2606.24530 · Hugging Face Daily Papers · 2026-06-24 · Source ↗

Frontis AI's NatureBench distils 90 tasks from peer-reviewed Nature-family papers across six scientific domains. The strongest evaluated model beats published state of the art on 17.8 percent of tasks. The companion NatureGym pipeline turns a paper into a containerised task package automatically.

NatureBench, the second-most-upvoted Hugging Face paper on June 24 at 48 upvotes, is a cross-discipline benchmark of 90 tasks distilled from peer-reviewed publications in the Nature family of journals, spanning six scientific domains. The construction pipeline, called NatureGym, takes a published paper as input, extracts the dataset and held-out test set, builds a containerised task package with an automated evaluator, and scores any candidate model against the source paper's reported state of the art. The framing question Frontis AI puts on the cover is whether coding agents can move beyond reproduction toward discovery. The answer is mostly no.

The strongest model the authors evaluated surpasses published SOTA on only 17.8 percent of the 90 tasks under the g greater than 0.1 criterion, the threshold the authors set for declaring an agent's solution meaningfully better than the published baseline. The methodological analysis pushes harder. When the authors trace the pathway by which agents succeed on the 17.8 percent of tasks they do solve, the bulk of the wins come not from genuine scientific invention but from methodological translation, the agent reframing the published problem as a familiar supervised-prediction task and applying a strong off-the-shelf model. The agents are not doing science, they are doing pattern matching from a science context onto the ML context where they have leverage.

The release ships the benchmark, NatureGym, and a public leaderboard with maintainer-side reproduction. The framing matters for the broader debate on AI-for-science. Recent OpenAI work on GPT-5 hypothesis generation, the announcement that an immunologist used GPT-5 Pro to predict an unpublished T-cell experiment, and the John Jumper move from DeepMind to Anthropic last week have all pushed the narrative that frontier models are starting to do real scientific work. NatureBench is the first systematic dataset that lets that claim be measured against the published SOTA in mainstream-journal science, and the first read is that the gap to actual discovery is wider than the demos suggest.

TechTrend

MWC Shanghai opens with humanoid robots on stage as GSMA declares the mobile mission shifted from connecting people to enabling intelligent machines

TechTimes · GSMA · 2026-06-24 · Source ↗

Director General Vivek Badrinath opened the keynote with humanoid robots walking the stage. China Telecom signed a Robotics-as-a-Service deal with AgiBot, the Shanghai company that crossed 10,000 humanoid robots shipped in March and led 2025 humanoid shipments at 39 percent global share per Omdia.

MWC Shanghai opened on Wednesday morning at the Shanghai New International Expo Centre with what the GSMA described as the first telecom keynote in which humanoid robots walked the stage before any human speaker. Director General Vivek Badrinath used the opening address to formalise a shift in the GSMA's pitch that has been building through the 2026 Mobile World Congress cycle: the mobile industry's mission, he said, is no longer to connect people, it is to enable intelligent machines. The framing is a deliberate echo of the 5G-Advanced standardisation push, which is now positioned by the GSMA as the nervous system for embodied AI, humanoid robots, drone fleets and autonomous vehicles.

The most concrete deal of the morning was a Robotics-as-a-Service partnership between China Telecom and AgiBot, the Shanghai-based humanoid robotics company that crossed its 10,000th humanoid shipment in March 2026 and ranked first globally in humanoid robot shipments for full-year 2025 with a 39 percent global share according to Omdia. The China Telecom deal puts AgiBot units into aviation and logistics workflows under a leasing model rather than a direct purchase, lowering the deployment threshold for the carrier's enterprise customers and giving AgiBot the recurring revenue line that hardware-only humanoid programmes have struggled to lock in. The Mobile AI Innovation Frontiers Zone on the show floor debuted the Humanoid Robot Football Penalties Challenge, an event-design experiment that explicitly bars pre-programming, hidden teleoperators and resets. The point of the rules is to force real-time autonomous decision-making, the gap in current humanoid software that the lab-demo videos rarely surface.

The broader exhibitor context, including announcements from Huawei, China Mobile, ZTE and a long tail of Chinese AI infrastructure vendors, sits across the next three days. The day-one signal is the framing. The GSMA, the consortium of mobile operators, is committing the standards roadmap and the carrier marketing budget to embodied AI as the next demand driver. The Robotics-as-a-Service model is the commercial pattern those operators are betting on to monetise it.

Tech

OpenThoughts-Agent ships 15,000 SFT traces and 720 RL tasks for training small agentic models, post-trains Qwen3-8B for Terminal-Bench and SWE-Bench

OpenThoughts · Hugging Face Daily Papers · 2026-06-24 · Source ↗

A Stanford, Berkeley, UT Austin, NYU, UW, UCSD, ASU, CMU, UCLA, UNC, TUM and LAION collaboration releases SFT and RL pipelines, NL2Bash plus InferredBugs source tasks, and OpenThinker-Agent-v1 post-trained from Qwen3-8B for terminal and SWE workflows.

OpenThoughts-Agent, the 27-upvote project on Wednesday's Hugging Face Daily Papers, is the open-source counterpart to the proprietary data pipelines that frontier labs use to train coding and computer-use agents. The release is a research collaboration anchored at Stanford, UC Berkeley, UT Austin, NYU, the University of Washington, UCSD, Arizona State, Carnegie Mellon, UCLA, UNC Chapel Hill, the Technical University of Munich and the LAION research network. The deliverable has three layers: a documented data-curation recipe for both supervised fine-tuning and reinforcement learning, the OpenThoughts-Agent-v1-SFT dataset of approximately 15,000 traces drawn from NL2Bash synthetic tasks and the InferredBugs corpus of bugs in C# and Java, and the OpenThoughts-Agent-v1-RL dataset of roughly 720 verified-task instances drawn from a curated subset of NL2Bash.

The first reference model trained on the recipe is OpenThinker-Agent-v1, post-trained from Qwen3-8B in two stages. Stage one is supervised fine-tuning on the 15,000 SFT traces. Stage two is reinforcement learning with task verifiers on the 720 RL instances. The targeted evaluation benchmarks are Terminal-Bench 2.0 and SWE-Bench, the canonical evaluations for terminal-agent and software-engineering-agent capability. The open-source community has been waiting for a published recipe that does for agentic post-training what OpenThoughts-1, the same group's prior release, did for reasoning post-training: provide enough data and enough method detail that a small model team can reproduce results without negotiating data licences. The release closes that gap for the terminal-agent surface specifically, and sets the template for follow-up datasets in browser-use, OS-use and software-engineering domains.

The reason it matters for the broader open-source-models picture is that the bottleneck for small open agentic models has stopped being base-model quality. Qwen3-8B, Llama 4 family, GLM-5.x and DeepSeek-V3 are all credible starting points. The bottleneck is post-training data. OpenThoughts-Agent puts a credible, openly published dataset and pipeline into the hands of the teams who want to ship 8 to 30 billion-parameter agent models without paying for a synthetic-data generation contract.

Tech

Kwai's MobileForge ships annotation-free adaptation for mobile GUI agents via hierarchical feedback-guided policy optimisation

Hugging Face Daily Papers · 2026-06-24 · Source ↗

MobileForge from kwai, at 34 upvotes on the day, removes the human-annotation step from cross-app mobile GUI adaptation by replacing it with a hierarchical feedback signal that drives policy optimisation on the agent's own rollouts.

MobileForge, the third-most-upvoted Hugging Face paper on June 24 from the kwai research team, attacks the annotation bottleneck that has slowed the cross-app generalisation of mobile GUI agents. The standard pipeline for training a mobile agent on a new app or a new device family today requires human annotators to label which UI element the agent should click, in what order and with what expected effect. The cost of that annotation is the main reason mobile agents work reasonably well on the apps inside their training distribution and degrade quickly on apps outside it. MobileForge replaces the human label with a hierarchical feedback signal computed from the agent's own rollouts. A top-level evaluator scores whether the agent completed the user-level task. A mid-level evaluator scores whether intermediate sub-goals were reached. A bottom-level evaluator scores individual action quality. The three layers feed into a policy-optimisation loop that improves the agent without per-step human supervision.

The result is a deployable agent that adapts to a new app from rollouts alone. The companion paper, also from kwai and at 33 upvotes, is MemGUI-Agent. Where MobileForge solves the no-annotation problem, MemGUI-Agent solves the cross-session memory problem: it manages context proactively so the agent can continue a multi-step task across app launches and time gaps. The two together describe a working stack for autonomous mobile workflows that does not depend on offline labelling. Both papers fit into kwai's broader Doubao-adjacent research push on consumer-facing agents, and both ship with code released to the project's GitHub.

Tech

MemGUI-Bench finds existing mobile agents use 5 to 11 percent of memory-relevant tasks and exposes five distinct failure modes

arXiv 2602.06075 · Hugging Face Daily Papers · 2026-06-24 · Source ↗

The kwai MemGUI benchmark covers 128 tasks across 26 mobile apps, 89.8 percent of which require cross-temporal or cross-spatial memory retention. The companion MemGUI-Eval pipeline runs progressive scrutiny with seven hierarchical metrics across 11 state-of-the-art agent systems.

MemGUI-Bench, the kwai paper at 33 upvotes on Wednesday's Daily Papers, opens with a finding that reframes the mobile-agent evaluation landscape. Existing mobile GUI agent benchmarks systematically under-test memory: between 5.2 and 11.8 percent of tasks across the major benchmarks actually require an agent to remember anything between session boundaries, and none evaluate cross-session learning explicitly. The MemGUI-Bench suite corrects the gap with 128 tasks across 26 applications, 89.8 percent of which require cross-temporal or cross-spatial memory retention to complete. The benchmark adopts a systematic memory taxonomy, evaluating 11 distinct agent systems across 5 architectural families.

The companion evaluation tool, MemGUI-Eval, runs the agents through Progressive Scrutiny, a staged LLM-as-judge protocol with seven hierarchical metrics. The headline result is that all 11 evaluated state-of-the-art agents show significant memory deficits relative to the task floor. Five distinct failure modes emerge: forgetting earlier user preferences, inability to resume after an app reload, contamination from prior sessions, loss of intermediate computed state, and failure to use memory for confirmation prompts. Five corresponding design implications are synthesised. The benchmark, evaluator and a leaderboard are open-source at the lgy0404/MemGUI-Bench repository on GitHub.

Combined with MobileForge's annotation-free adaptation, the two kwai releases on June 24 amount to a deliberate research push: kwai is staking out the mobile-agent surface as research territory that is undersupplied by the US labs, and is publishing the gap and the method together.

Tech

Tsinghua AIR ships AOHP, an open-source OS-level agent harness for personalised, efficient and secure interaction

arXiv · Hugging Face Daily Papers · 2026-06-24 · Source ↗

The Institute for AI Industry Research at Tsinghua publishes AOHP at 25 upvotes, an OS-integrated agent harness that runs personalised workflows under per-user permission scopes and lower-level system controls than browser-only agents can reach.

AOHP from the Institute for AI Industry Research at Tsinghua is an open-source operating-system-level agent harness. The distinction matters. The mainstream agent-frameworks landscape has been split into two camps: browser-controlled agents like the OpenAI computer-use family or Anthropic's Claude Computer Use, which act on whatever a user has open in a browser tab, and CLI agents like Claude Code or OpenAI Codex, which act through a shell. Neither has clean access to the user's persistent application context. AOHP sits beneath both. It is an OS-level harness that lets the agent see what application is in focus, what files are open, what processes are running, and which devices and peripherals are connected, with each capability mediated through a permission scope and an audit trail.

The framing pitch in the paper is three-way: personalisation, efficiency and security. Personalisation comes from the OS-level state, the agent can adapt to what the user actually does instead of what they tell it. Efficiency comes from skipping the round-trip through a browser DOM or a terminal subprocess. Security comes from the explicit per-capability permission model. The release is open source and is positioned as the building block for downstream agentic-OS work, the same architectural direction Microsoft, Apple, Anthropic and Google have all hinted at in 2026 without yet shipping an open reference implementation. AOHP is the first such reference from a major Chinese academic lab.

Tech

EventVLA puts event-driven visual evidence memory into long-horizon vision-language-action robot policies

arXiv · Hugging Face Daily Papers · 2026-06-24 · Source ↗

Shanghai AI Lab's EventVLA at 17 upvotes incorporates event-camera-style evidence memory into long-horizon vision-language-action policies, the canonical bottleneck in robotic manipulation pipelines that fail on multi-minute tasks.

EventVLA from Shanghai AI Lab tackles the long-horizon problem in vision-language-action models for robots. Standard VLAs, the architectural class that includes Google's RT-2, Physical Intelligence's pi-zero and pi-0.5, and the various Open-X-Embodiment-trained models, work well on tasks with a short horizon and full visual context. They lose ground on tasks that take minutes, require remembering what was done two steps ago, or require returning to a previously visited surface in the scene. EventVLA introduces an event-driven visual evidence memory: instead of carrying the full vision token stream through the entire task, the model maintains a compressed event log that fires when something visually salient happens, then queries that log when later reasoning needs the earlier context. The architectural inspiration is event cameras, the hardware-side neuromorphic-vision sensor that already operates on a fire-when-changed signal pattern, but the implementation here is entirely software-side: it works against standard RGB streams from conventional cameras, with the event mechanism applied to the token representations.

The performance claim in the paper is that EventVLA closes a meaningful share of the long-horizon gap on multi-task evaluation suites, particularly on tasks where memory of an earlier scene state is decision-relevant. The lab released the work alongside a companion robotic-manipulation paper on world value models from ByteDance Seed and the InSight self-guided VLA skill-acquisition paper from Stanford, both also in Wednesday's Hugging Face list. The cluster signals where the robotics research community is focusing this summer: not on the next generic VLA size bump, but on the substructures that make multi-minute robot autonomy reliable.

Tech

SAILING Lab at CMU and MBZUAI publishes a Critique of Agent Model, naming structural limits in current autonomous-agent architectures

Hugging Face Daily Papers · 2026-06-24 · Source ↗

The Critique of Agent Model paper at 10 upvotes from SAILING Lab argues current agent architectures are bottlenecked by structural assumptions that conflate planning, execution and reflection, and proposes architectural changes that separate the three.

SAILING Lab, the joint research unit of Carnegie Mellon University and MBZUAI in Abu Dhabi, published Critique of Agent Model on Wednesday, June 24 at 10 upvotes. The paper sits outside the day's benchmark-and-recipe pattern and takes the structural argument seriously: current autonomous-agent architectures, the authors argue, are bottlenecked not by the size of the underlying language model but by structural assumptions about how planning, execution and reflection are wired into the same model call. The Critique catalogs the assumption set that the agent-framework community has converged on without formal scrutiny: single-model-as-planner-and-executor, one chain of thought per task, self-reflection through prompted critique, and tool-use as a special-case generation pattern. Each of these, the paper shows, has documented failure modes that are not addressed by scaling the underlying model.

The proposed architectural direction separates planning, execution and reflection into distinct components with explicit interfaces, similar to the agentic-RL community's move toward separate policy and value networks. The paper is theoretical rather than empirical, but the structural argument lines up with the failure analyses in MemGUI-Bench (same day, kwai), NatureBench (same day, Frontis AI) and the broader literature on long-horizon agent failures. The paper is worth reading specifically because it is one of very few pieces of research-community criticism aimed at the agentic-architecture consensus rather than at any single model.

ApplicationEconomic

Uber caps employee AI-coding-tools spend at $1,500 a month after exhausting its 2026 AI budget in four months

Simon Willison · 2026-06-24 · Source ↗

Uber introduces a $1,500 monthly per-tool spending cap on AI coding assistants including Claude Code, Codex and Copilot after the company's annual AI budget ran dry in four months on unexpected developer adoption.

Uber introduced a $1,500 per-employee monthly cap on AI coding-tool spend on Wednesday, June 24, following internal disclosures that the company exhausted its 2026 AI budget within the first four months of the fiscal year. Simon Willison flagged the policy on his weblog, framing it as a reasonable response to unexpected demand for agentic coding software rather than a backlash against the tools. The cap applies per tool, which is the part of the policy that signals the operational maturity of the decision: Uber is not restricting which tools developers can use, only how much per-developer-per-tool spend can flow without management sign-off.

The economics behind the cap are the same economics every enterprise rolling out Claude Code, Codex, Copilot or Cursor at scale has run into in 2026. Token costs for agentic coding workflows are an order of magnitude higher than the chat-style coding assistance budgets that companies sized in 2025. A single Claude Code session with sub-agent orchestration through a moderate-complexity refactor can burn through tens of dollars of inference; a heavy week of agentic-mode usage from one developer can comfortably exceed $1,000 of token spend. Multiply across a 5,000-developer engineering organisation and a 12-month budget set last fall and the burn rate that Uber surfaced is unsurprising. The Anthropic Claude Design 2.0 launch in mid-June introduced shared token budgets at the admin level precisely to address this category of failure mode, and the Vercel Connect launch added per-environment short-lived credentials that make budget enforcement at the integration boundary tractable.

The signal value of the Uber policy is that the AI-coding-tools cost line has moved from an experimental, growth-with-it budget category into a managed engineering-tools cost centre with caps and exception processes. The customer-side maturity is starting to catch up to the rate at which the tools themselves ship new spending surfaces.

Application

Microsoft's third annual AI in Education report finds 92 percent of students and leaders on AI, ships Copilot Notebooks and a Study and Learn Agent at no extra cost

Microsoft Source · 2026-06-24 · Source ↗

Methodology: 3,345 K-12 and higher-ed respondents across the United States, United Kingdom, Australia, Brazil, Japan and Saudi Arabia. 88 percent of educators on AI, 77 percent of students and 53 percent of educators report no formal AI training, 87 percent of educators view AI fluency as important to student futures.

Microsoft published the third edition of its annual AI in Education report on Wednesday, June 24, alongside the launch of a new set of Copilot tools for Microsoft 365 Education. The headline finding is uptake. 92 percent of students and education leaders report using AI for school-related purposes, with 88 percent of educators on it. 78 percent of leaders, 76 percent of educators and 65 percent of students say their AI use for school has increased over the past year. The methodology is 3,345 respondents drawn from K-12 and higher-ed institutions across the United States, the United Kingdom, Australia, Brazil, Japan and Saudi Arabia.

The skills-gap finding is the part the report leans on for the product launch. 77 percent of students and 53 percent of educators say they have not received formal AI training, even though 87 percent of educators and 79 percent of students agree that knowing how to use AI well is important for student futures. 66 percent of educators and 52 percent of students want institutional AI training on a monthly or quarterly cadence. Academic integrity is the leading concern, named by 41 percent of students and 42 percent of educators.

The product set Microsoft shipped alongside the report addresses the gap inside the existing M365 footprint. Unit Plans in Teach surfaces standards-aligned lesson planning through Copilot. Student AI Guidelines and Learning Groups in Assignments give teachers a responsible-AI scaffolding to attach to any assignment. Learning Zone is a Windows 11 in-class experience with educator controls. Copilot Notebooks turns class materials into an AI-powered structured study guide with self-testing built in. The Study and Learn Agent inside Copilot Chat is a tutoring agent built for research-based learning guidance rather than answer-supply. All of these are included at no additional cost for A3 and A5 license holders. Matt Jubelirer, Microsoft's general manager for education, framed the rollout in tooling terms rather than rhetoric.

"Educators around the world are embracing AI as a classroom ally, and they're now asking not if, but how to make the most of it."— Matt Jubelirer, General Manager, Microsoft Education
TechApplication

DOE unveils Aires Tide, the first AI-designed flight test vehicle built under the Genesis Mission

HPCwire · NNSA · 2026-06-24 · Source ↗

The Department of Energy's National Nuclear Security Administration announces Aires Tide, a proof-of-concept flight test vehicle developed using AI-driven design optimisation, high-performance computing and additive manufacturing under the Genesis Mission.

The Department of Energy's National Nuclear Security Administration unveiled Aires Tide on Wednesday, June 24, a proof-of-concept flight test vehicle developed using artificial intelligence, high-performance computing and additive manufacturing under the Genesis Mission. The Genesis Mission is the DOE programme launched earlier in 2026 to apply frontier AI to national-security engineering challenges including weapons-stockpile stewardship, materials design and aerospace systems. Aires Tide is the programme's first publicly disclosed deliverable that went past simulation to hardware.

The methodology that produced the vehicle is the part that matters for the broader AI-for-engineering picture. The design loop ran inside DOE's HPC infrastructure, with AI optimisation iterating across geometry, materials, manufacturability and aerodynamic-performance criteria simultaneously rather than sequentially. The resulting design is then printed directly through additive manufacturing, the route that lets the AI-suggested geometries actually exist as parts. Conventional aerospace design loops cannot reach those geometries because they assume subtractive machining or composite layup as the manufacturing step. The combination of AI-driven design optimisation and additive manufacturing is the operational story of the announcement: not the vehicle itself, but the design-to-hardware pipeline that produced it.

The geopolitical context is the DOE's expanding role in frontier-AI deployment. The same Genesis Mission lent compute to Reflection AI for the open-weight model work that closed a $6.3 billion SpaceX Colossus deal on Monday, and the NNSA is participating in the broader US-China competition framing that drove the Trump administration's June 2 executive order on AI innovation and security. Aires Tide is the visible end of that policy push.

Policy

FCA's Nikhil Rathi tells the City legislation will never keep up with AI and reframes the regulator as a stewardship body

PYMNTS · Financial Conduct Authority · 2026-06-24 · Source ↗

In a keynote at the Agents of Change conference titled Re-thinking regulation for the age of AI, the CEO announces the FCA will explore agentic AI as a wholesale-market first responder, publish a Mills Review on AI in retail finance and an AI practice guide later in 2026, and run a Supercharged Sandbox built with Nvidia and Google.

Nikhil Rathi, the chief executive of the UK Financial Conduct Authority, used a Wednesday keynote at the Agents of Change: Generative and Agentic AI in Financial Services 2026 conference to formalise a strategic pivot in how the FCA operates against AI risk. The speech, titled Re-thinking regulation for the age of AI, argued that traditional rule-making cannot keep pace with the rate at which AI capability is being deployed across financial services. Rathi was explicit in the framing, with three quoted lines that the City press carried in full.

"Technology is moving much faster than many regulatory paradigms. Legislation will never keep up."— Nikhil Rathi, CEO, Financial Conduct Authority
"A growing part of our role will be stewardship, as well as supervision. In some areas, we will still need detailed rules. But in others, traditional rule-making simply won't work."— Nikhil Rathi, Re-thinking regulation for the age of AI, June 24

The announced workstreams operationalise the framing. The FCA will explore agentic AI as a first-responder tool for wholesale market monitoring, the regulator's surveillance function over trading venues and large-volume order flow. The Mills Review on how AI could reshape retail financial services publishes in the coming weeks. A separate AI practice-guidance document, naming good and poor AI practices for regulated firms, lands later in 2026. The FCA is also expanding its sandbox programme into what it calls a Supercharged Sandbox, run in partnership with Nvidia and Google to give participating firms access to frontier-scale compute for regulated-experiment workloads.

The structural reading is that the FCA is the first major financial regulator to commit publicly to a stewardship-plus-rule-making hybrid as its operating model for AI oversight. The European Securities and Markets Authority, the US Securities and Exchange Commission and the Hong Kong Monetary Authority have all signalled discomfort with the pace of AI deployment in finance over the past two quarters. The FCA is the first to attach an operating doctrine and a published workplan to the discomfort.

TechTrend

Polymarket abandons the GPT-5.6 launch window as June 22 to 28 odds collapse from 83 percent to 18 percent and July prices at 94 percent

Yahoo Finance · Polymarket · 2026-06-24 · Source ↗

Traders priced an 83 percent probability earlier in the month that OpenAI would ship GPT-5.6 between June 22 and June 28, on $960,325 in contract volume. With the window closing and no model live, the same contract collapsed to 18 percent yes and 83 percent no. July is now the favoured outcome at 94 percent by month end.

Polymarket's GPT-5.6 release-date market collapsed through Wednesday afternoon as the June 22 to 28 launch window narrowed without an announcement. Earlier in June the market had priced an 83 percent probability that GPT-5.6 would ship inside the six-day window. The volume backing that price was substantial, $960,325 in contract value as of June 15 according to the market book. By Wednesday's close the same yes-side contract was trading at 18 percent, with the no-side at 83 percent. The market for a July release, the next-out window, repriced upward through the same period and is now sitting near 94 percent by end of July.

OpenAI's GPT-5.6 family had been the focus of testing-channel speculation through June. TestingCatalog flagged GPT-5.6 and GPT-5.6-Pro in internal A-B testing on June 18 and pointed to a launch window starting the week of June 22. Capability rumours include a context window expansion from 1 million tokens in GPT-5.5 to approximately 1.5 million tokens, plus a companion GPT-Bidi-1 bidirectional voice model with High, Medium and Instant tiers. The pricing rumour is aggressive, with reports suggesting OpenAI would extend its existing roughly-half-of-Anthropic token-pricing position with further per-token cuts to widen the gap.

The market signal is what it is: traders have removed the assumption that the model ships this week. The collapse does not mean the model is not coming, only that the timing rumour the market was paid into was wrong. The most likely operational explanation is the Trusted Access expansion that OpenAI shipped on Wednesday for GPT-5.5-Cyber. Bringing a new flagship online in the same week as a major safety-tier expansion of the prior flagship is operationally hard, and OpenAI's release cadence tends to space those announcements by at least one week.

Tech

llama.cpp tags build b9782 the morning of June 24 with the usual cross-backend binary fan-out

GitHub ggml-org/llama.cpp · 2026-06-24 · Source ↗

The reference open-source LLM inference engine cuts build b9782 with binary distributions for Ubuntu, macOS and Android, plus CUDA, ROCm, SYCL and OpenVINO GPU backends. The release continues a daily cadence that has run uninterrupted through June.

The llama.cpp project tagged build b9782 on the morning of Wednesday, June 24. The release ships the project's standard cross-platform binary set: Ubuntu Linux builds for CPU and the GPU backends, macOS builds for Apple Silicon and Intel, Android builds for ARM, and the GPU-backend variants covering CUDA for Nvidia, ROCm for AMD, SYCL for Intel and OpenVINO for the Intel deep-learning runtime. The release cadence in June has been daily or near-daily, with the project absorbing fixes from contributors against the steady drip of new open-weight model architectures coming out of Qwen, DeepSeek, Mistral, Moonshot and the Llama derivatives.

llama.cpp's quiet operational reliability is the part that matters for the open-source ecosystem. While the spotlight moves between vLLM, Ollama, MLX and SGLang on the inference-engine landscape, llama.cpp remains the default substrate that the others link against or wrap, and the project's GGUF format remains the dominant weights-on-disk format for local inference. A daily release cadence with the cross-backend matrix kept clean is the kind of operational discipline that does not draw attention until it stops, and it has not stopped through 2026.