Mistral Large 4 Weights Land, Cut Cost Per Task 63%
By @sharedot · · 7 pages
- AI Frontier
- Open Weight Models
- Mistral
Mistral released weights for its trillion-parameter Large 4 model, cutting cost per benchmark task from $5.54 to $2.06.
The weights finally arrived
Mistral released downloadable open weights for its trillion-parameter Large 4 model on October 29 under a custom license, after telling customers on October 6 it would publish them "by the end of the month" — an earlier October 27 date reported by The Next Web and The New Stack was not met, and ActuIA notes a Hugging Face listing still showed no files as of October 28.
According to VentureBeat, the release adds roughly three weeks of reinforcement learning since the API preview, and the model remains a sparse mixture with 1.05 trillion total parameters and 52 billion active parameters, 1 million-token context, and vision-plus-text input.
Artificial Analysis scores the released checkpoint at 26 on Intelligence Index v5, one point above the preview.
The real news is efficiency, not capability
Independent tests, per VentureBeat, show Large 4's biggest advances over the preview are efficiency, lower cost and fewer fabricated answers, with only a one-point capability gain on Artificial Analysis's Intelligence Index v5 (26 versus 25).
The economic shift is stark: cost per Intelligence Index task fell 63%, from $5.54 for the preview to $2.06 at the high reasoning setting, close behind Muse Spark Open's $2.02.
On the Coding Agent Index, Artificial Analysis ranks ML4 second among open-weight entries at 48, just one point behind GLM-5.3, while using roughly 8.5 million tokens per task — about 47% less than GLM's 15.9 million.
The catch: DeepSeek V4.1 Flash 1017 still scores higher at just $0.60 per task, and MiMo-V2.6-Pro at $0.26.
The license is not plain MIT
VentureBeat reports the Ecosystem Open-Weight License is a modified MIT license with a material restriction: Section 2 requires a separate commercial agreement if the licensee or an affiliate runs a model-as-a-service business — offering third parties inference or post-training with meaningful control over inputs, parameters or training data.
Crucially, there is no revenue threshold; the trigger is business activity, not company size. Internal use and end-user products whose model capabilities are solely embedded in specific features are exempt, as are services that merely relay requests to models hosted elsewhere.
Mistral chief revenue officer Marjorie Janiewicz told VentureBeat customers want to "own" their model, customize it and keep business continuity without depending on a provider's ongoing permission.
The benchmark scoreboard is shifting underfoot
ActuIA's analysis adds context on how scores are read: Artificial Analysis's Intelligence Index v5 moved 60% of scoring behind private test sets, dropping every listed open model by 12 to 16 points, so Large 4's 26 sits on the new scale where the best published weights trail Claude Opus 5.5 by far more than before.
Benchmark context also matters for Mistral's coding claims: VentureBeat notes Mistral self-reports 41% on Terminal-Bench 4.0 while Artificial Analysis's Intelligence Index evaluation reports 25% on that same test, and Mistral's 19.7% FrontierCode 2.0 lead over DeepSeek in its own chart reverses in AA's assessment, where ML4 sits at 15% behind DeepSeek's 29% and GLM-5.3's 22%.
Mistral also reports a Surge Code blind preference study where ML4 wins 64% of comparisons against GLM-5.3 and Kimi K3.
The open-weights geopolitics behind the release
The timing lands amid a live policy debate the Manhattan Institute highlights: its scholars Arpit Gupta, Brandon Fuller and Sean Speer argue Chinese labs like Qwen and DeepSeek currently dominate open-weight AI, and propose an antitrust safe-harbor trade in which American frontier labs would be obligated to release capable open-weight models on a predictable schedule after independent testing.
Mistral's release is exactly the kind of voluntary, one-off opening that memo says developers cannot yet build on with confidence from US labs.
What comes next for builders: Large 4 is downloadable on Hugging Face, servable with vLLM or SGLang, and available through Mistral Studio, OpenRouter, Microsoft Foundry with private Azure deployment, and Nvidia NIM endpoints, with three reasoning settings — low, medium and high — letting teams reserve deeper reasoning for the hardest requests.
Sources
- venturebeat.com › Mistral Large 4 arrives as the top open model outside China with improved coding performance and 63% lower cost-per-task
- actuia.com › Mistral Large 4 drops from 38 to 25 on new Artificial Analysis index
- manhattan.institute › New Policy Memo: An American Answer to China's Open AI Models