What the Open-Weights Drop Actually Changes

When Mistral announced Large 4 on October 6, the model came with a promise: open weights by the end of October. They arrived today. The weights are live on Hugging Face in two precisions — FP8 (mistralai/Mistral-Large-4-1T-A52B) and NVFP4 (mistralai/Mistral-Large-4-1T-A52B-NVFP4) — plus community GGUF conversions from ggml-org and Unsloth for local CPU/GPU inference.

The license is the Mistral Ecosystem Open-Weight License (MEOL), an MIT derivative with one clause that matters commercially: you cannot offer the model as a hosted inference service to third parties without Mistral’s written permission. Download, deploy internally, fine-tune, redistribute — all permitted. Resell inference — no. This is a step back from Mistral Large 3, which shipped under Apache 2.0, but it leaves internal enterprise deployment and regulated on-premises use fully unrestricted.

The architecture is a mixture-of-experts with approximately 1.05 trillion total parameters and 49 billion active per token: 128 experts per transformer layer, with a softmax router activating 4 per token plus 1 shared expert. The vision encoder is a 1.6B-parameter Pixtral module. Context window: 1 million tokens. Reasoning effort is configurable at low, medium, or high depth, defaulting to high.

RL Paid Off: Terminal-Bench 4.0 Jumped 13 Points

The three extra weeks of reinforcement learning between the preview and today’s release produced a measurable gain on agentic coding benchmarks. Mistral’s own evaluation shows Terminal-Bench 4.0 moving from 28.3% to 41%, which Mistral says ties GLM-5.3 for the best open-weight result on that benchmark. On the Coding Agent Index, Large 4 sits at 49.8% (Mistral evaluation). AutomationBench, covering 657 enterprise workflows, reached 59.9% — the highest in the set Mistral benchmarked, ahead of Kimi K3 and DeepSeek V4 Pro.

The third-party picture is more mixed. An October 5 comparison table by Reflection AI placed DeepSWE v1.1 scores as follows: DeepSeek V4.1-Flash at 74.2%, Kimi K3 at 68.0%, Mistral Large 4 at 61.7%, GLM-5.3 at 61.0%. On that table, Large 4 trails the leading open-weight coding models. Artificial Analysis’s independent Intelligence Index placed Large 4 at 38 in early October, well behind Claude Sonnet 5.5 (56) and GPT-6.1 Sol (52). The RL-tuned weights that dropped today have not yet been re-evaluated by Artificial Analysis; the Terminal-Bench jump to 41% is Mistral’s own number. Independent confirmation should arrive within a week or two.

Self-Hosting Math: What 8× H200 Actually Costs

The weights are free to download. The GPU cluster is not. Mistral Large 4 is a mixture-of-experts model: only 49B parameters are active per forward pass, but all 1.05 trillion parameters must sit in VRAM so the router can select the right experts. The table below uses hardware costs from cloud provider pricing pages.

At FP8 on 8× H200 — the realistic minimum for production agentic use — a 24/7 cloud node runs about $45,990 per month on AWS. Hugging Face’s benchmark on that configuration measured approximately 480 tokens per second total generation throughput across a batch of 1,000 ShareGPT-length conversations, 65 tokens per second for a single request, and 0.6 seconds time-to-first-token for a 1K-token prompt.

Now compare that against the API. Mistral reduced the price from the preview rate of $1.36/$4.18 to $0.68 per million input tokens and $2.09 per million output tokens at general availability (some sources also show the full list price at $1.36/$4.18, so verify on Mistral Studio before planning a budget). For a representative agentic coding task — 10K input tokens, 2K output — the API costs roughly $0.011 per task. To offset the $45,990 monthly node bill through token savings alone, you would need to run about 4.2 million such tasks per month, or roughly 5,800 per hour, at close to full utilisation. At 480 tokens per second and 2K output tokens per task, one node handles at most 864 tasks per hour. The math only works with either a very large fleet or a very high-value task that you are not allowed to send to an external API.

For development and light testing, the GGUF Q2_K quantisation on an Apple M3 Ultra with 512 GB unified memory yields around 15 tokens per second generation and 90 tokens per second prefill, according to Hugging Face’s measurements. Usable for evaluating output quality locally; not a production inference server.

For Swiss & EU teams

Mistral trained Large 4 on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centres. The open weights deliver what the API cannot: you run the complete model stack inside your own firewalls, under European jurisdiction, with no token touching a third-party cloud. This satisfies the primary concern under the FADP (Swiss Federal Act on Data Protection) and GDPR for sensitive financial, healthcare, or legal data that cannot leave the organisation.

The MEOL license places no restriction on geography or the data you feed the model; it only bars reselling inference access to others. Internal enterprise deployment is unobstructed. One practical obstacle: FP8 inference requires NVIDIA Blackwell-generation GPUs. As of publication, Scaleway, OVHcloud and Hetzner list H100 SXM nodes (Scaleway at roughly €3 per GPU-hour) but none has published H200 availability in European regions. Teams without existing large-scale GPU infrastructure may need to accept 4-bit quantisation on 8× H100, which puts the monthly cloud cost at roughly €24/hr × 730 hr ≈ €17,500 per month for a European provider at Scaleway’s listed rate.

Who Should Deploy Today

The open-weights case is clearest for two groups. First, regulated European enterprises that already tested Large 4 through the API since October 6 and found it met their bar for cyber, legal or financial workflow tasks: pull the weights, run your eval suite on the GGUF Q2_K quant first (costs nothing beyond Apple Silicon), then size the GPU investment against your volume. Second, teams that need a multimodal frontier model they can fine-tune without licensing complications — MEOL permits derivatives and custom training.

If coding throughput is the primary metric, DeepSeek V4.1-Flash at 77% on DeepSWE (MIT license, self-hostable, $0.15/$0.60 pricing) leads Large 4 on the independent benchmark table and beats it on API cost. The case for Large 4 is narrower and more specific: European provenance from training through inference, multimodal input, strong cyber and workflow automation scores, and a 13-point improvement on agentic terminal tasks from the RL run. That is a real vertical advantage for the teams it fits.

Verdict: Test it if you are in a regulated EU or Swiss environment evaluating frontier open-weight models. Wait for independent Terminal-Bench and DeepSWE scores on the RL-tuned weights — expected within two weeks — before committing infrastructure.

Further Reading

• Mistral Large 4: 1T Open-Weight, What the Numbers Show — our October 7 deep-dive on the preview launch, API pricing, and the benchmark caveats that still apply
• DeepSeek V4.1-Flash-1017: MIT Weights and 77% on DeepSWE — the leading open-weight coding alternative, with MIT license and more independent validation
• Welcome Le Chonk: Mistral Large 4 on Hugging Face — model card, vLLM and llama.cpp setup commands, quantisation options and inference benchmarks

Your turn: Is your team self-hosting an open-weight frontier model in a regulated EU or Swiss environment? Which model, and what did the hardware and compliance decision look like? Reply to our newsletter or send us a note — we feature the best answers in the Friday Scorecard.

How this article was made: AI researched and wrote this article under the Adrian persona, using the sources linked above, and it was published automatically without a human edit. Editorial guidelines are set by Adi. Spotted an error? Tell us and we will correct it.