mistral-large-4
Meant to excel at
-
Mistral's largest model: about 1T MoE (52B active), natively multimodal, 160+ languages, trained and served in European datacenters.
Vendor's own site mistral.ai -
Pitched for coding and agentic work (DeepSWE v1.1 61.7%), cyber defense (Cybench 93%) and visual grounding on documents and charts.
Vendor's own site mistral.ai -
Handles spreadsheets, slides and PDFs; reports resisting 93.3% of prompt-injection attacks on Lakera's B3.
Vendor's own site mistral.ai -
Best end-to-end cybersecurity result (CyberGym-E2E about 82%) and top intelligence among models from outside the US and China.
Independent evaluator tomshardware.com -
Fast output (about 116 tokens/s, 1.46s first token); won most blind human comparisons versus GLM-5.3 and Kimi K3.
Independent evaluator tomshardware.com
Watch out for
-
Hallucinated on 41.9% of knowledge questions and answered only 25.8% correctly; do not rely on it answering from memory.
Independent evaluator eesel.ai -
Token-heavy: about 200M output tokens to run the index versus an 81M median; about $1.13 per task at list price.
Independent evaluator tomshardware.com -
Reliability gaps in agentic coding: sometimes ended turns with no output; multi-file builds struggled.
Practitioner report mindstudio.ai -
Still a public preview: weights unreleased, specs conflict across pages, and Terminal-Bench (26.8%) and HLE (35.0%) are weak.
Independent evaluator gradually.ai -
Current API prices are a preview sale at half list ($0.68/$2.09 vs $1.36/$4.18); the docs give no end date.
Vendor's own site docs.mistral.ai
Benchmarks
Agentic coding percentile 0.09
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| deepswe v1.1 | 61.7 | 0.00 | Vendor claim, relayed implicator.ai | 2026-10-06 |
| terminal-bench 4.0 | 28.3 | 0.19 | Independent evaluator aireleasetracker.com | 2026-10-06 |
Overall percentile 0.17
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| artificial analysis intelligence index | 38 | 0.17 | Independent evaluator artificialanalysis.ai | 2026-10-19 |
Percentile is the share of tracked models this model beats on that benchmark, averaged across the group. Only benchmarks at least four tracked models share are scored.
Practitioner reports
- +
Plotly's data-analytics test: Large 4 passed 32 of 43 questions for $1.94 per run vs 25 of 43 for Mistral Medium 3.5 at $20.51; misses were judgment errors such as reporting a 7-day total as an average and missing hidden duplicate rows.
Task: data-analysis | Compared with mistral-medium-3.5 | 2026-10-09
Practitioner report developersdigest.tech - -
Practitioners see it as a solid mid-priced option, not a coding leader: Surge AI blind ratings put it behind Claude Opus 5, and it is verbose (about 200M output tokens for the AA index vs 81M median).
Task: feature | Compared with claude-opus-5 | 2026-10-09
Practitioner report developersdigest.tech - -
A developer with a custom-language benchmark found Large 4 roughly level with Kimi K3 and a Claude Sonnet model but much slower than the Sonnet model and pricier than both; a Vals Index comment scored it below GLM-5.3 at higher cost.
Task: research | Compared with kimi-k3 | 2026-10-09
Practitioner report developersdigest.tech - +/-
Scored 42/80 on eight agentic coding tasks: perfect on a local fine-tune and web page, but silent on two tasks and weak at multi-file coordination.
2026-10-08
Practitioner report mindstudio.ai - +
A Plotly engineer says accuracy on their data-analytics benchmark rose from 58% to 74%; not on the Pareto frontier but good enough for data analytics. (HN comment relayed by eesel.)
Compared with mistral-large-3 undated
Practitioner report eesel.ai - -
A commenter says it would need to be significantly better than GLM 5.3-flash to justify its price. (HN comment relayed by eesel.)
Compared with glm-5.3-flash undated
Practitioner report eesel.ai - -
A LocalLLaMA user found it dated for code analysis, image classification and chess, far behind the models they normally use. (relayed by eesel.)
undated
Practitioner report eesel.ai - +
A singularity user argues it is probably good enough for most corporate workflows and values its European origin for data sovereignty. (relayed by eesel.)
undated
Practitioner report eesel.ai