Moonshot pushed Kimi K4 Base to Hugging Face at approximately 09:00 UTC+8 on Sunday, a 2.6-trillion-parameter mixture-of-experts with 150 billion active per token, 10-million-token native context, Apache 2.0 license. The release lands 91.4 on Terminal-Bench 2.1, the top of the open-weight board. Hosted inference prices at $0.80 input and $3.20 output per million tokens on Moonshot's own API with a 90 percent cached-input discount; day-zero support arrived inside four hours across vLLM v0.34, llama.cpp b11503, SGLang Router v2.3, Hugging Face Transformers v5.22 and Ollama 0.6.1. The model is the third major open-weight frontier release of the week and the capstone of a 72-hour sequence that reset the open-source leaderboard three times.
Reflection AI opened Beam on Thursday morning Pacific as a 501-billion-parameter MoE with 23 billion active, Apache 2.0 license, trained from new weights over four weeks on 10,500 Nvidia GB300 GPUs with more than 100 million rollouts and roughly 1.3 billion sandbox evaluations peaking at 883,006 vCPUs across 23 clusters on Google Cloud and AWS. The release pairs the weights with an Apache 2.0 open operating system, Mirror CLI (Rust and Python), one-command Kubernetes deployment, and 33 named launch partners; Dell and HPE install on customer hardware, Baseten serves through OpenRouter from day one. Beam reaches about 80.1 on Terminal-Bench 2.1 using roughly 25,000 tokens per attempt against GLM-5.2's 80,000-token path to the same band. Reflection sets no hosted price; partners price their own serving.
Meta Superintelligence Labs released Muse Spark Open on Thursday morning Pacific as an 816-billion-parameter MoE with 42 billion active, 1-million-token context, placing weights on Hugging Face, Amazon SageMaker, Google Cloud Model Garden and Microsoft Foundry Models with day-one hosted access through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The Muse Model License keeps commercial use free for most developers but requires a separate agreement from any hoster past $20 million in annual gross revenue, prohibits military and critical-infrastructure uses, and asks downstream products to display a "Built with Muse Spark Open" label. Meta's own benchmark table has Muse Spark Open at 61.8 on JobBench against Thinking Machines Lab's Inkling at 31.1 and NVIDIA Nemotron 3 Ultra at 27.2, at 89.2 on MCP-Atlas ahead of every rival listed, and at 76.4 on Terminal-Bench 2.1 behind Kimi K3 at 85 and both GLM 5.3 and Qwen3.8.
"The weights are open, the operating system is open, and the whole evaluation fleet that trained it is documented. The lean path on Terminal-Bench is the thing we are proudest of."— Misha Laskin, Reflection AI CEO, Beam launch post
Why the 72-hour sequence lands as a single story
The three releases are not independent. The capability gap between Kimi K4 at 91.4 and Reflection Beam at 80.1 is the gap between a 2.6-trillion-parameter capstone and a 501-billion-parameter lean release; the gap between Beam and Muse Spark Open at 76.4 Terminal-Bench is architectural rather than scale. The common structural fact is Apache 2.0 or near-Apache licensing across all three, day-zero support across six inference stacks inside four to six hours, and hosted-price listings that make even the lowest-priced hosted path optional. The pattern repeats in the second-tier releases of the week: Mistral opened Codestral 3 as a 48B dense coding model under the Mistral Research License at 71.8 on SWE-Bench Lite, NVIDIA opened Nemotron 3.5 Super VL at 120B total and 12B active under OpenMDW with 600-plus output tokens per second on Artificial Analysis agentic tasks, Black Forest Labs opened FLUX 3 Klein at 5.6B with native 4K generation, IBM released Granite Switch 4.1 bundling Decision, Guardian and Granite-Speech adapters as Activated LoRAs over 8B and 30B bases that share the base KV cache.
"Open weights at 2.6 trillion parameters with a ten-million-token context under Apache 2.0 is not a milestone, it is a floor. The question for the hosted-API margin is now attestation and safety RL rather than capability."— Clem Delangue, Hugging Face CEO, Sunday X post
Where the frontier labs sit at the end of the week
Microsoft completed pre-training on MAI-Base-2, a 2.8-trillion-parameter MoE with 80 billion active, 16,000 GB200 GPUs over 38 days with 50 trillion tokens, claimed 4.2x to 12.3x pre-training FLOPs efficiency against DeepSeek-V4-Pro-Base, no distillation. Weights will ship through Foundry to trusted partners rather than open release. OpenAI shipped GPT-6 Astra Mini with vision on every tier including the free plan at $0.20 input and $0.80 output per million tokens with the vision surcharge removed. The pattern across the three closed releases from Microsoft, OpenAI and Anthropic (whose Opus 5.6 Mythos stayed unchanged at $18/$90 cyber-tier) is that the proprietary band has not matched the open-weight burst on either sticker price or weight availability. The Vercel AI Gateway Production Index for October put open-weight share of hosted inference requests at 42 percent against 31 percent in September, with Anthropic below 50 percent for the first time at 43.9 percent. Hugging Face published its October State of Open AI report putting open-weight share of hosted-inference requests at 42 percent against 31 percent in September, with the Reflection and Muse releases expected to extend that number into November.