Mistral Large 4 Open Weights: Vendor TB 41% vs AA's 25%, High Default Effort, and $0.70/$2.10 Before You Self-Host

Attributed compile + one small offline run on my box

Primary source: Mistral Large 4 (Le Chonk) is now open weight by Mistral AI, published 2026-10-29 (weights after ~3 weeks of extra RL since the early-October public preview).

Secondary: Welcome Le Chonk: Mistral Large 4 on Hugging Face by Aritra Roy Gosthipaty (Hugging Face, 2026-10-29); VentureBeat — Carl Franzen for the Artificial Analysis vs Mistral Terminal-Bench / FrontierCode split and token-per-task notes (live VB fetch returned HTTP 429 on this draft day — numbers below follow Mistral’s post + the VB figures already confirmed for this series; do not treat a blocked fetch as missing coverage).

Model id / card: mistralai/Mistral-Large-4-1T-A52B (also FP8 / NVFP4 / GGUF siblings).

Benchmarks and prices below are vendor-reported (or clearly attributed to HF / VentureBeat’s reading of Artificial Analysis) unless I say otherwise. This box did not call chat completions and did not download weights. Offline preflight = public HTTP scrapes + list-price arithmetic only.

---

Mistral dropped open weights for Mistral Large 4 (“Le Chonk”) on 2026-10-29: a ~1T-parameter natively multimodal MoE that had already been in public preview earlier in October. The OSS-day story is not “a new model family name.” It is more RL since preview, a reasoning-effort knob, a list price, and a license with a MaaS commercial gate. For builders, the useful move is: separate Mistral’s charts from Artificial Analysis’s own evals, set effort deliberately (default is high), and decide API vs self-host before you swap a production model string.

Repo naming on Hugging Face is Mistral-Large-4-1T-A52B. The model card frames roughly ~1T total / ~52B active (granular MoE: 128 routed experts, 4 active + 1 shared). An HF blog line once said 49B active — prefer the A52B / card naming when you write configs. Context is 1M tokens on the card. Architecture (short): MLA + DSA-style sparse attention, Pixtral vision encoder — enough to know it is sparse MoE + long context, not enough to skip a golden eval.

Do not merge vendor benches with Artificial Analysis numbers

Mistral’s OSS post leads with coding agent claims. Label every row as who measured it.

Mistral-reported (OSS post, 2026-10-29)

Claim (Mistral / named partner)	Figure	Notes
Cognition FrontierCode 2.0 (private)	19.7% lead (vendor)	End-to-end “would a maintainer merge?” framing
Artificial Analysis Coding Agent Index	48%, effectively tied with GLM-5.3	Mistral also claims ~47% fewer tokens/task vs that peer on average
Terminal-Bench 4.0	41% (+13 pp vs preview)	Mistral says lead among open-weight in their chart; ~30% fewer output tokens vs preview
AA Cyber Index	51%	Defensive find / reproduce / patch framing (vendor)
Malicious cyber refusal (Mistral measure)	avg 96.6%	JailbreakBench / StrongREJECT / AgentHarm cyber prompts — their open-model comparison
Finch	73.3%	Finance / spreadsheet workflows (vendor)
BrowseComp	84.5%	From an SFT checkpoint Mistral cites at 75%
B3 agent security attack resistance	97.1% (preview 93.3%)	Prompt-injection / agent-attack resistance (vendor)

VentureBeat’s roundup is useful because it puts Mistral’s numbers next to Artificial Analysis Intelligence Index rows without pretending they are the same harness:

Metric	Mistral claim	AA Intelligence Index eval (per VentureBeat)
Terminal-Bench 4.0	41%	25% (ML4 high)
FrontierCode	19.7%	15%

Same model name, different eval contexts. Do not paste “41% TB4” into a slide that cites AA’s Intelligence Index table, and do not call AA’s 25% a Mistral walk-back. HF’s launch blog still surfaces some preview-era coding rows (e.g. Terminal-Bench 28.3%, Coding Agent Index ~49.8% in that post’s table). Treat those as not the OSS-day Terminal-Bench claim of 41% after more RL. When in doubt: cite the date + source, or re-run your own tasks.

Token efficiency (Mistral + VB reading): on Coding Agent Index work, Mistral positions ML4 near GLM-5.3 on score while using far fewer tokens — VB cites roughly ~8.5M tokens/task vs GLM ~15.9M (~47% less). That is cost/latency leverage, not permission to skip quality gates.

Reasoning modes: low / medium / high — and none still means high

OSS-day ML4 exposes three reasoning modes (Mistral table):

Mode	Mistral’s recommended use
Low	Simple tasks; replaces the old None mode
Medium	Everyday tasks
High (default)	Hardest / high-precision work

Hugging Face’s Transformers / Inference Endpoints notes match the trap: reasoning_effort accepts low | medium | high, default high, and none still maps to high — reasoning is not truly off. If your migration code still sends none from an older Mistral preview client, you are not getting a cheap no-think path; you are paying high by default.

Builder rule: set effort explicitly per route. Use low for classification / short edits, medium for normal agent turns, high only where your eval shows it earns the tokens. Leaving the default untouched is a silent bill and latency multiplier.

List price: $0.70 in / $0.07 cached / $2.10 out

Mistral says that starting today (OSS announce day) API list price is $0.70 input / $0.07 cached input / $2.10 output per 1M tokens. Hosted surfaces named in the post: Mistral Studio, Mistral API, OpenRouter, plus NVIDIA NIM / hosted endpoints and Microsoft Foundry. Self-host path: Hugging Face weights + vLLM or SGLang (FP8 and NVFP4 official; GGUFs from Unsloth / llama.cpp).

List-price arithmetic (my assumptions — not a bill)

Offline preflight used those three rates only. Assumptions: one trajectory = 80k input + 20k output; “80% cache hit” means 80% of input tokens bill at the cached rate; no tax, retries, or tool-overhead tokens.

Scenario	USD (list math)
1 traj, no cache	$0.0980
1 traj, 80% input cache	$0.0577
40-turn loop, no cache	$3.9200
40-turn loop, 80% input cache	$2.3072
1M in + 1M out, no cache	$2.8000

Cache is the swing factor on long agent loops. Measure real hit rates before you put the middle rows in a forecast. Effort=high will also inflate output tokens relative to low — price the effort knob and the rate card together.

License: open weights ≠ free MaaS

License is Mistral Ecosystem Open-Weight (modified MIT). Mistral’s summary: free to use / modify / fine-tune / commercialize, with one hard condition — businesses that operate a Model-as-a-Service offering need a separate commercial license before commercial MaaS use. Self-hosting inside your product is not the same as renting the weights out as a public model API. Read the LICENSE on the HF card before you put ML4 behind a multi-tenant endpoint you sell.

Separately: Mistral mentions a specialized cybersecurity variant for vetted design partners (e.g. Orange Cyberdefense) under contractual safeguards. That is access-gated, not the public HF checkpoint. Do not assume the public OSS weights are that partner build.

Checklist before you swap models (API / OpenRouter vs HF + vLLM/SGLang)

1. Freeze a golden eval on your agent failures (repo patches, terminal tasks, tool traces). Vendor TB 41% and AA’s 25% are both someone else’s harness.
2. Pin the model id — start from mistralai/Mistral-Large-4-1T-A52B (or the NVFP4 / FP8 sibling you actually serve). Do not guess strings. Docs also surface API-style ids such as mistral-large-4-0; confirm against the surface you bill on.
3. Set reasoning_effort explicitly. Default is high. Treat none as still high.
4. Price both columns: list $0.70/$2.10/$0.07 and expected output tokens at your chosen effort. Re-run the arithmetic with your real prompt sizes.
5. Self-host only if you have GPU budget (HF / partners discuss multi-GPU / Blackwell-class NVFP4 paths; VB notes quantized serving on 4× B200/B300 class setups — treat as secondary hardware rumor until you size it yourself), ops for vLLM/SGLang, and a license read for any MaaS-shaped product.
6. API / OpenRouter first when you need a same-day A/B without multi-GPU bring-up — then decide whether weight ownership is worth the ops.
7. Cyber workloads: use the public model under your own policy and sandbox. Partner cyber variants are gated; public refusal rates are Mistral’s measure, not your compliance program.
8. Do not conflate Mistral charts, HF blog tables, and AA Intelligence Index rows in one “leaderboard” screenshot.

What I actually ran (offline)

/workspace/reputation/runs/30-mistral-large-4/ — 2026-10-30 America/Toronto:

MISTRAL_API_KEY: NOT SET
OPENROUTER_API_KEY: NOT SET
Chat completions: NOT ATTEMPTED

HTTP 200: mistral.ai/news/mistral-large-4-oss/
HTTP 200: huggingface.co/mistralai/Mistral-Large-4-1T-A52B
HTTP 200: huggingface.co/blog/mistral-large4
HTTP 200: mistral.ai/pricing ; docs.mistral.ai models page
Needle hits on OSS page: $0.70 / $2.10 / $0.07, TB 41%, reasoning modes, MaaS license language
Model card: A52B id + FP8/NVFP4 siblings present

List-price math (assumptions above):
  1 traj no cache $0.0980 | 80% cache $0.0577
  40-turn no cache $3.9200 | 80% cache $2.3072
  1M+1M $2.8000

NOT tested: live chat; weight download; vLLM/SGLang; effort behavior;
Terminal-Bench / Coding Agent re-score; real cache hits; partner cyber variant

Costs are arithmetic on Mistral’s stated list rates under assumptions I state. They are not a bill, not latency, and not quality.

When ML4 might become a default

Only after a callable path (API or your own serve) and a golden eval win on your tickets:

1. Coding / terminal / cyber-defense workloads where your current open or mid-frontier default fails.
2. You have chosen effort per route, not left high everywhere.
3. Cost model survives realistic cache and high-effort output inflation.
4. License review is done if anything you ship looks like MaaS.
5. You can roll back without rewriting orchestration.

Weights on Hugging Face are the story this week. Treat Mistral’s 41% and AA’s 25% as two labeled columns, set effort on purpose, and run the $0.70/$2.10 math on your trajectories before you call the swap done.

---

YongBo Yu · Toronto AI engineer · https://yongbo-yu.vercel.app · GitHub YongBoYu1