Gemini 4 Omega landed on Hugging Face at 06:12 UTC-7 Tuesday, with the Google DeepMind release post going live on deepmind.google at the same minute. The model is a sparse mixture-of-experts with 2.1 trillion total parameters, 128 billion active per token, 2,048 experts in each of 96 MoE layers, top-32 routing, Multi-Query Latent Attention 2 across 160 heads, and a native 2-million-token context window held at 512,000 tokens per request for public inference. Training ran on 24,576 TPU v6e chips across two data centers for 56 trillion tokens, with a 4.1 trillion-token math-and-code mid-training phase and a 900-billion-token reasoning stage. The weights, tokenizer, and the training recipe for the mid-training and reasoning phases are open under a new Gemini Research + Commercial License, which replaces the Research-only cap applied to earlier Gemma releases and permits commercial use by any operator with annual revenue below $10 billion.
What the license covers and what it does not
The Gemini Research + Commercial License (GRCL v1.0) is the first license from Google that permits general commercial use of frontier Gemini weights. The $10 billion revenue cap applies to the operator serving inference, not to downstream users; a startup serving a Fortune 100 customer is permitted. The license forbids weight redistribution to any party on the US Entity List, requires a 90-day advance notice for any derivative work released publicly, and reserves takedown rights for Google in response to NIST-level safety findings. The Apache 2.0 question raised by Qwen 5 Max's Sunday release remains open; Google product lead Sissie Hsiao said in the release post that "GRCL is calibrated to the regulatory environment we actually operate inside, which includes export controls and the EU AI Act." Hugging Face confirmed day-zero mirroring with licensing banners in the model card.
Benchmarks and the frontier band
Gemini 4 Omega sits inside the frontier band on four of five published benchmarks. On DeepSWE v1.1 it posts 83.7, against Opus 5.6 Mythos at 83.5, GPT-6.1 Sol at 82.4, DeepSeek V5 Thinking at 83.4, and Qwen 5 Max at 81.2. On OSWorld 2.0 it scores 76.2, closing to Gemini 4 Argon Thinking's 76.1 and Opus 5.5's 75.4. MMLU-Pro at 95.1 edges Qwen 5 Max's 94.8 and Opus 5.5's 94.6. CWE-bench defender hit 73 percent against Mythos at 70.3 and Qwen 5 Max at 72. ARC-AGI-2 at 62.1 is the top open-weights score; closed-weights Opus 5.6 Mythos holds 64.3. The fifth benchmark, ExploitBench, Google held at 9.2 percent and documented the restriction in the model card as a filtered-dataset result rather than a capability floor, matching the pattern Alibaba used with Qwen 5 Max on Sunday.
"Open weights at the frontier is a two-move game. Alibaba made the first move Sunday. Google making the second move Tuesday means every lab now has to answer whether the hosted price is a convenience fee or a floor." — Dylan Patel, SemiAnalysis note to clients
The hosted price and the self-host arithmetic
Vertex prices Gemini 4 Omega at $1.80 input and $9 output per million tokens with a 95 percent cached-input discount for prompts over 32,000 tokens. DeepSeek V5 Thinking is $0.14 / $0.56 under MIT; Qwen 5 Max is $0.60 / $2.40 under Apache 2.0; Mistral Large 4 Thinking GA announced Tuesday at €0.75 / €3 under MRL; GPT-6.1 Sol is $1 / $5; Opus 5.6 Mythos is $18 / $90. The open weights make the Vertex price optional for operators running at scale: at 2.1T parameters with 128B active, a FP6 Blackwell SM100 deployment runs an estimated $0.42 input and $2.10 output per million tokens on a 32-GPU Rubin GB300 preview rack, inside the DeepSeek V5 band once hardware costs amortize. SemiAnalysis published its "Open frontier at $0" analysis Tuesday afternoon pricing the self-host economics at scale, concluding that the hosted-API margin for frontier-band inference compressed to under 60 percent for the first time this quarter.
Day-zero stack support
vLLM shipped v0.33 within four hours of the drop with Gemini 4 Omega top-32 expert routing, native support for the 2,048-expert routing table on H200 and Blackwell, and a Jev-compatible envelope output API matching the one Hugging Face Transformers v5.20 shipped Sunday. llama.cpp cycled builds b11465 through b11470 across Tuesday morning Pacific, landing FP8, FP6 and FP4 quantizations and committing the first production Rubin GB300 kernels inside the same release — kernels that were preparation-only in b11445 Sunday. SGLang shipped Router v2.2 with sparse-MoE caching at 2.3x throughput on Blackwell. Hugging Face Transformers v5.21 added Gemini 4 Omega day-zero with the output_type="structured" API extended to a new output_type="jev-envelope" alias for direct TypeSafe compatibility. Ollama 0.6.0 shipped multi-model orchestration that routes between Gemini 4 Omega, Qwen 5 Max, DeepSeek V5 and local Mistral variants based on a tiered policy engine. MLX added Apple-silicon support in a 10:38 AM UTC-7 commit, running Gemini 4 Omega 4-bit on 192GB M4 Ultra at 51 tokens per second.
"We built Omega to be used by labs that are not us. The license, the recipe, the mid-training data pipeline, the SFT mix. The $10B cap exists because of regulatory asymmetry, not strategy. We will remove it when the regulatory environment permits." — Demis Hassabis, Google DeepMind CEO, release post
What came before
Google's open-weights history runs from the Gemma 1 and 2 releases of 2024, through Gemma 3 at the 70-billion-parameter scale in 2025, to the EmbeddingGemma 2 release Thursday October 8 at 740 million parameters inside AI Edge Foresight. Gemini 4 Omega is the first time Google has opened weights for a frontier-tier flagship, closing a public gap with Meta's Spark (Spark Max 2 Research 1.3T previewed Sunday), Alibaba's Qwen (Qwen 5 Max 1.1T Sunday), and DeepSeek (V5 820B MIT). The move follows a16z's "System One + System Two" Sunday essay predicting 60 percent of agent inference calls routes through structured-decision models by Q2 2027, and SemiAnalysis's Monday note arguing that the sub-$1 output floor meant hosted-API margins would compress to below 50 percent inside two quarters. Hassabis said during the DeepMind all-hands Tuesday that Omega "closes the open-weights gap and opens a conversation about where our commercial moat actually sits." Internal Google slides on Vertex-plus-Agent-Gateway integration, previewed Thursday last week, point to the moat sitting at the agentic routing and Workspace integration layer rather than the model weights.
Caveats and open questions
The model card documents several limitations. The 2-million-token context ships at 512,000 tokens per request for public inference because of a KV-cache sharding limit the research team expects to resolve in a December point release. The Rubin GB300 kernels in llama.cpp b11470 are first-production, not optimized; a benchmark run against Nvidia's internal results shows a 1.6x throughput gap closing to within 5 percent in the b11475 draft slated for next week. The GRCL $10 billion revenue cap means Microsoft, Amazon, Meta, Apple, OpenAI, Anthropic, Alibaba, Tencent, Baidu and ByteDance cannot host Omega inference commercially without a separate agreement; several of these parties were approached in the two weeks before launch under NDA. OpenAI declined comment; Microsoft confirmed it would use Vertex-hosted Omega inside the Copilot pooled routing announced Tuesday rather than self-host.
"The GRCL cap is the Vertex business model dressed as a license. It says: if you're not us or them, you're welcome; if you're them, call us." — Nathan Benaich, Air Street Capital, State of AI daily note
What this means for the price war
Three frontier-band open-weights releases inside six days — DeepSeek V5 Thinking Thursday October 8, Qwen 5 Max Sunday October 11, Gemini 4 Omega Tuesday October 13 — has repriced the floor of the market. Hosted frontier inference pricing: $0.14 / $0.56 (V5), $0.60 / $2.40 (Qwen 5 Max), $1.80 / $9 (Omega), $1 / $5 (Sol), €0.75 / €3 (Mistral Large 4 Thinking GA), $12 / $60 (Mythos Lite), $18 / $90 (Mythos). The weight-level cost of self-hosting all three open frontier models continues to fall as Rubin GB300 broad sampling moves to November 11 confirmed Tuesday. The closed-weights frontier ceiling now has to defend a 60-to-100x price premium over open-weights alternatives that match or exceed benchmarks on four of five results. OpenAI's response is expected inside the week; a Mythos Pro repricing from Anthropic was previewed in enterprise briefings Monday but not announced.