Mistral Large 4's Final Weights Land: Same Brains, 63% Cheaper, Fewer Hallucinations

Mistral just shipped the downloadable open weights for Large 4 (ML4), its trillion-parameter model, and the independent numbers tell a more interesting story than "new model beats old model." The biggest gains aren't in raw intelligence. They're in cost, efficiency, and how often the thing makes stuff up.

Artificial Analysis ran its October 29 assessment and gave ML4 at high reasoning a 26 on Intelligence Index v5, up from 25 for the preview. That's one point above Muse Spark Open, enough to make ML4 the top open-weight model built outside China on that particular leaderboard. It's still six points behind the actual open-weight leaders, DeepSeek V4.1 Flash 1017 and GLM-5.3, both running at max reasoning. So: incremental intelligence gain, meaningful positioning claim, not a frontier leap.

The economics moved a lot more than the benchmark score. AA reports ML4 now costs $2.06 per Intelligence Index task, down from $5.54 for the preview. That's a 63% reduction, and it brings ML4's cost close to Muse Spark Open's $2.02. DeepSeek and MiMo still undercut it substantially ($0.60 and $0.26 respectively), scoring higher on intelligence while costing less. ML4's efficient output-token usage doesn't automatically make it the cheapest option on the table; input traffic, caching, and token rates all factor into the final bill.

The API pricing backs this up: $0.70 per million input tokens, $0.07 for cached input, $2.10 for output, roughly half the preview's $1.36/$0.14/$4.18. AA's assessment confirms these are the real, measured rates for the shipped checkpoint, not a projection.

What Actually Changed Since the Preview

The model's core identity didn't change. It's still a sparse mixture-of-experts system, 1.05 trillion total parameters, 52 billion active, 1 million token context window, trained on roughly 3,800 Nvidia Grace Blackwell GPUs (revised down from the initially reported ~4,000). What changed is three more weeks of reinforcement learning and feedback from customers and security partners before the weights actually went public.

A couple of spec discrepancies from the original October 6 coverage are worth clearing up. The active parameter count moved from 49 billion to 52 billion between reports, but Mistral's Hugging Face listing explains this isn't model growth: 49B describes per-token activity, while 52B also counts embeddings and output layers. Different measurement, same model.

On the coding side, AA's Coding Agent Index puts ML4 at 48, one point behind GLM-5.3's 49 and two ahead of DeepSeek's 46, using Mistral's own Vibe agent software. The efficiency story here is real: ML4 used 8.5 million tokens per task versus GLM's 15.9 million, a 47% reduction. Mistral calls the scores "effectively tied." AA ranks it second. Both can be true; a one-point gap isn't statistically meaningful on its own.

Where things get murkier is Mistral's other coding claims. The company's self-reported Terminal-Bench 4.0 number is 41%, up from 28% for the preview. AA's own Intelligence Index testing of the same model puts Terminal-Bench at 25%. Same model, different harness, different number. Similarly, Mistral's FrontierCode 2.0 chart shows ML4 at 19.7%, edging out DeepSeek's 19.3%. AA's independent run has ML4 at 15%, trailing DeepSeek's 29% and GLM-5.3's 22%. The lesson for anyone citing these benchmarks: check which harness produced the number before you compare across vendors.

Mistral also ran a blind preference study through Surge Code, with ML4 winning 64% of head-to-head comparisons against GLM-5.3 and Kimi K3, 56% against DeepSeek V4.1 Flash, and 22% against Claude Opus 5.5. That's human preference data, a genuinely different kind of signal than task-completion scores, and shouldn't be conflated with either.

The hallucination numbers are arguably the most practically useful result here. On AA-Omniscience v1.1, ML4's hallucination rate dropped to 28%, down 13 points from the preview's 41%, matching Muse Spark Open's rate. But accuracy sits at 26%, well below Muse's 42% and DeepSeek's 40%. Translation: ML4 got better at not making things up, but that's not the same as getting better at being right. For enterprise teams, a model that stays quiet when unsure is easier to supervise, but it still needs to answer enough questions correctly to be useful.

Output-token efficiency improved sharply too: 82,000 output tokens per Intelligence Index task versus the preview's 135,000, a 39% cut. The seven open-weight models scoring above ML4 all burn between 113,000 and 225,000 tokens per task, so ML4 lands in a genuinely good capability-to-cost zone even if it's not the absolute cheapest.

On licensing, the Mistral Ecosystem Open-Weight License is a modified MIT license with one real catch: there's no revenue threshold, but Section 2 requires a separate commercial agreement if the licensee or any affiliate runs a "model-as-a-service" business, meaning giving third parties API or hosted access with meaningful control over inputs, parameters, or training data. Internal use is exempt. End-user products with model capabilities embedded solely within specific features are exempt too. Pure request-relaying services are exempt. Everything else involving hosted inference or post-training for third parties needs that separate license.

"I don't want to get disconnected. I want to own my model," said Marjorie Janiewicz, Mistral's chief revenue officer, describing the pitch to customers in an interview with VentureBeat. "Second, I want to customize it to ensure that I derive value. And then third, it's business continuity."

Key Takeaways

ML4's story isn't a capability breakthrough, it's a maturation story. The intelligence score moved one point. The cost per task dropped 63%. The hallucination rate fell 13 points. Coding efficiency improved by nearly half the token count for comparable performance. For enterprises evaluating whether to self-host a near-frontier open-weight model, those operational numbers matter more than the headline Intelligence Index score, especially since DeepSeek and GLM still lead on raw capability and sometimes on cost too. The real decision point isn't "is ML4 the smartest open model," it's "does ML4's balance of efficiency, licensing terms, and ownership fit what I'm building," and that answer depends heavily on your own workload, not just a leaderboard.