The phrase "49 billion active parameters" has been doing a lot of work in the coverage of Mistral Large 4. It implies something modest-a mid-range model you might run on a decent server. The full weight file will be around 525 GB at 4-bit quantization, before KV cache and runtime overhead. That is not a server; that is a small cluster.
Mistral launched a public preview of Mistral Large 4 on October 6, 2026: a 1-trillion-parameter model with 49 billion active parameters, text-and-image input, and open weights promised by the end of the month. The API is live. The weights are not yet available; Mistral plans to release them on October 27. And there is a detail buried in the announcement that most coverage glossed over: the reinforcement learning phase is still ongoing, and according to Mistral, the improvements have not yet plateaued-meaning the preview's benchmark results are preliminary and may still change before the weights are made available.
That matters for any team trying to evaluate ML4 today. What you are calling through the API is not the model whose files you will eventually download.
What "49B active" actually means for your hardware bill
An MoE model routes each token through a small subset of experts, so only a fraction of parameters fire for any given forward pass. Active parameters drive inference compute-ML4 processes tokens at roughly the cost of a 50B dense model-but total parameters determine VRAM requirements, because the full weight matrix lives in memory regardless of how much fires per forward pass.
The word "active" describes computation, not the whole cost of owning or serving the model. It does not mean that the complete model has the memory footprint of a 49B dense model.
The math is straightforward once you run it:
Estimated planning figures are about 630 GB of GPU memory for a 4-bit checkpoint, 1.26 TB for FP8, and 2.5 TB for BF16.
| Precision | Weight size | GPU config needed |
|---|---|---|
| BF16 | ~2.1 TB | Multiple 8-GPU nodes |
| FP8 | ~1.05 TB | 8× H200 (just fits, 78 GB spare) |
| 4-bit | ~525 GB | 8× H100 (on paper, 115 GB spare) |
At FP8, a single 8-GPU server with H200s has about 78 GB to spare after weights. At 4-bit on H100s, there's about 115 GB left on paper. That spare headroom evaporates quickly when you add KV cache for long-context requests and concurrent users. Mistral's own pages disagree slightly on the active parameter count-49 billion in the announcement and 52B in the docs-so treat it as about 50B until the model card settles it.
The benchmark numbers you should actually use
The reinforcement learning phase, in which the model improves through trial and error and feedback, is still ongoing. Mistral reports that gains have not begun to slow. The benchmark results are preliminary, and the company expects them to change.
RL is the post-training stage where the model improves by being scored on its own attempts at tasks, and Mistral says its run currently uses about 3,000 GPUs, generates roughly 33 billion tokens a day, and "is showing no signs of saturation."
So what do the current numbers actually say? From Mistral's own announcement, confirmed against the Artificial Analysis and vals.ai evals that Mistral cites:
ML4 scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4. Its combined Coding Agent Index score of 49.8% places it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max.
In a blind human evaluation with Surge AI, professional annotators rated model outputs on a 1-5 scale, with model identities hidden. ML4 Preview ranked second of five models at 3.74, ahead of Kimi K3 (3.59) and GLM-5.3 (3.60), and behind only Claude Opus 5 (4.22).
The independent aggregate tells a different story. Artificial Analysis scores Large 4 Preview at 38.4 on its Intelligence Index-a large jump over Mistral Medium 3.5 (14 points) and Large 3 (9 points)-but among open models, it currently ranks eighth.
The honest reading: ML4 is genuinely competitive in cybersecurity and agentic business workflows, and genuinely mid-pack on the composite intelligence leaderboard. The cybersecurity story is genuinely differentiated: a model that scores 82% on reproduce-and-patch while remaining open-weight and self-hostable is unusual.
General-purpose coding agents are a weaker fit right now. The Terminal-Bench score of 28.3% reflects real limits in long-horizon CLI tasks.
The API vs self-hosting cost crossover
Mistral Large 4 is callable today through a public-preview API at $1.36 per million input tokens and $4.18 per million output tokens, with open weights promised but not yet shipped.
Mistral launched at half of its published list price, and Mistral has not set an end date for that discount.
Budgeting at the $1.36 input / $4.18 output list rate is the safer long-term assumption, even while the lower sale rate is what actually gets billed today.
Now compute what self-hosting buys you. The minimum plausible configuration-eight H100 80 GB cards at 4-bit-currently rents for roughly $20-25/hour on major cloud GPU markets. That is around $15,000-18,000/month before storage, networking, engineering time, and the ops overhead of a model that needs distributed serving. At list API pricing, you would need to push roughly 100 million output tokens a month to reach that cost. That is a lot of throughput for most teams.
The one scenario where self-hosting wins immediately is data residency. Once the files are public, banks, ministries and manufacturers can run Large 4 on their own servers and keep data inside Europe. That is the actual pitch-not cost savings.
Beagle in action#legal-ops, 10:47amWhat to wait for before making a decision
There are three things that do not exist yet and that materially change the calculus:
The weights themselves. Due October 27. Mistral plans to publish architecture details, additional benchmarks, post-training methodology, and security test results alongside the weights drop.
The license. Mistral Large 3 shipped under Apache 2.0. Mistral's documentation lists Large 4 as "Open" without naming a licence, and VentureBeat reports the weights will arrive under a custom Mistral licence whose terms were not detailed. Until those terms are published, nobody can say whether commercial self-hosting will be unrestricted.
The final checkpoint. The RL run behind the preview is still training. The model you can call through the API today won't be the model whose weights ship on October 27.
A teammate like Beagle-sitting in Slack, calling whatever model your team has approved-doesn't need to solve the hosting question. But an engineering team evaluating whether to switch from a hosted frontier model to a self-hosted ML4 for a privacy-sensitive pipeline should wait thirteen more days before writing any infrastructure tickets.
Mistral Large 4 self-hosting: common questions
How much VRAM does Mistral Large 4 need to run?
The minimum practical footprint for self-hosting Mistral Large 4 is around 630 GB of GPU memory at 4-bit quantization-enough for weights plus basic overhead. FP8 precision requires roughly 1.26 TB. No single consumer or prosumer GPU comes close. Plan for an 8-GPU server at minimum, pending Mistral's official hardware guidance with the weights release.
Is Mistral Large 4 actually open source?
Not yet, and the final terms are uncertain. The preview API is live, but the weights are due October 27. Mistral Large 3 shipped under Apache 2.0; Large 4 is labeled "Open" but no license has been published. Until the license ships alongside the weights, commercial self-hosting terms are unconfirmed.
How does Mistral Large 4 compare to DeepSeek and GLM on coding?
On DeepSWE v1.1, ML4 scores 61.7%, ahead of DeepSeek V4 Pro (57%) but behind GLM-5.3 (61%). On the broader Artificial Analysis Intelligence Index, ML4 scores 38 against GLM-5.3 at 45 and Kimi K3 at 44. The cybersecurity evals are where ML4 leads among Western open-weight models.
Should teams use Mistral Large 4 today or wait?
The API is usable now for prototyping, especially for security-sensitive document workflows and business automation. For production commitments-self-hosting decisions, model selection, cost modeling-wait for the October 27 weights, license, and final checkpoint. The preview benchmarks are explicitly preliminary; the RL run is still live.
What makes the MoE architecture a self-hosting trap?
Mixture-of-Experts models are cheap to run per token but expensive to hold. All expert weights must live in GPU memory even though only a small fraction activate per token. "49 billion active parameters" describes inference cost, not storage cost. At 4-bit precision, ML4's 1.05 trillion total parameters still require around 525 GB of weight storage, not the ~25 GB you'd need for a true 49B dense model.