Skip to content
Today
Sign in

BuildNot yet confirmed elsewhere1 publisher2 min readPublished

Self-hosting Mistral Large 4 means loading 1.05 trillion parameters to run 52 billion per token

Mistral AI put Large 4, a 1.05-trillion-parameter mixture-of-experts model, into public preview on Oct. 6 and promised weights by October's end. Self-hosters can size memory now, though Mistral says the training run behind the preview is still going.

The Engineer · Build desk

How we use AISend a correction

Illustration accompanying Self-hosting Mistral Large 4 means loading 1.05 trillion parameters to run 52 billion per token
Generated illustration
Weights alone fill terabytes at two bytes each Weight memory by precision (released precision not stated), the context window in Mistral's documentation, and the vision encoder's size.

Weights alone take about 2.1 TB at two bytes per parameter and 525 GB at half a byte. Mistral's documentation lists a 1 million token context window; the preview API limits some third-party providers to 524,000 tokens. A separate vision encoder has 1.6 billion parameters.

Weights alone fill terabytes at two bytes each
FigureValueNoteClaim
Weights alone, two bytes per parameter, about2.1 TB525 GB at half a byte per parameter10
Context window, Mistral's documentation1,000,000 tokensSome providers' preview cap: 524,0005
Separate vision encoder1.6 billion params3

What happened

  • Only 52 billion of those parameters are active per token, and the router picks them from a larger pool of smaller experts than coarser layouts use.
  • Mistral trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers.
  • Cybersecurity organizations and government authorities are red-teaming the model before release, and vetted defenders will get a reduced-moderation version.
  • Artificial Analysis scores the preview 38 on its Intelligence Index with reasoning on, behind GLM-5.3 at 45 and Kimi K3 at 44 but first among open-weight models from outside China.

Why it matters

  • constraint Weights alone come to 1.05 TB at one byte per parameter and 2.1 TB at two, before any KV cache, so the 52 billion active figure cannot be used to size memory.
  • decision Hardware ordered now is sized against a preview checkpoint. Buyers choose between committing before the final weights, precision and licence are known, or waiting.
  • contradiction Mistral's documentation lists a 1 million token window, while the preview API caps some third-party providers at 524,000, so long-context behavior at the documented length cannot be tested there yet.

Active parameters are 4.95% of the total [9]. That share governs compute per token. In our view the weight memory still follows all 1.05 trillion [11], because routing is decided per token: a batch of concurrent requests sends tokens to many different experts, and the router can want any of them on the next step. Mistral says its granular layout improves throughput and task specialization [2]. Mistral nicknamed the model "Le Chonk" [1], and for memory the name is fair.

The scores are Mistral's, on Mistral's chosen workloads. It reports a 49.8% Combined Coding Agent Index, ahead of DeepSeek V4 Pro and Qwen3.8 Max [13]. For that to transfer, your agent tasks would have to resemble the ones behind DeepSWE v1.1 (61.7%) and Terminal-Bench 4 (28.3%) [14], and the checkpoint you run would have to match the preview. The second condition is the weaker one, because Mistral expects the model to keep improving before the weights ship [6].

In the blind human evaluation run with Surge AI, professional annotators rated the preview 3.74 out of 5. That put it second among five models, 0.48 points behind Claude Opus 5 at 4.22 [15][16].

The claim that Large 4 is the largest open-weight model from a European lab depends on weights that have not shipped. It comes from a single dev.to write-up of Mistral's announcement [12]. That write-up does not give the weight precision, layer count, attention layout or licence terms, and it does not say whether the 1.05 trillion includes the separate 1.6-billion-parameter vision encoder [3]. Without those, the KV cache at the documented context window cannot be sized.

What to watch

  • The weights release by end of October: shipped precision, licence terms, and whether it matches the final RL checkpoint.
  • Whether third-party providers lift the 524,000-token cap toward the documented 1 million.
  • Whether the red-teaming by cybersecurity organizations and government authorities delays the weights.

Clarity's read

What the record supports and how the coverage leans. The claims behind it follow.

Reality

Evidence40
Adoption
Insufficient
Hype gap+25
Incentives
Insufficient
Confidence50
Why these scores

Claim ledger

Ranked by verification strength, evidence, and original report placement.

  1. [1]

    Mistral AI released a public preview of Mistral Large 4 on October 6, 2026, a 1.05-trillion-parameter multimodal Mixture-of-Experts model that the company has nicknamed "Le Chonk."

    ReportedSupportedSource: dev.to write-up of Mistral's announcementView cited source
  2. [2]

    Mistral Large 4 has 52 billion active parameters per token; its granular MoE design has the router select from a larger pool of smaller experts instead of a few large blocks, which Mistral says improves both throughput and task specialization.

    ReportedSupportedSource: dev.to write-up; design benefit is Mistral's claimView cited source
  3. [3]

    A separate 1.6-billion-parameter vision encoder handles image inputs.

    ReportedSupportedSource: dev.to write-up of Mistral's announcementView cited source

Sources

1 independent publisher whose own reporting we read for this story.

  1. dev.to

    1 article · October 12, 2026

    Mistral Large 4: Inside Europe's First 1-Trillion-Parameter Open-Weight Model

Share your take

Let Clarity write the post for you.

Signed-in readers get a short post drafted on this story in the register they choose — narrative, analytical, or a direct position — editable to the last word before it goes anywhere. The share buttons at the top of this story work without an account.

Create a free account to composeSign in

Topics and entities

Follow any of these and your For You feed starts watching them — no settings page required.

Topics

Entities

Loading related stories