Skip to content
ENTITY PROFILE / KNOWLEDGE GRAPH

Mistral Large 4

RSS

Mistral Large 4, also called Le Chonk, is a trillion‑parameter open‑source AI model from Mistral. Its weights were released on 29 Oct 2026 and are available for download on Hugging Face and the Mistral API, allowing businesses and governments to deploy the model on‑premise. The model is marketed for sovereign cyber‑defence, granting organizations frontier intelligence that cannot be removed because the weights are open. It performs state‑of‑the‑art in cybersecurity and agentic coding tasks, ranking highest on Artificial Analysis’ Coding Agent Index. Compared to the Chinese competitor GLM‑5.3, it achieves comparable performance while using 47 % fewer tokens per task, making it more economical. Available in FP8 NVFP4 and GGUF variants, it is distributed under the Mistral Ecosystem Open‑Weight License and uses Pixtral as the vision encoder and LeRope for rotated position embeddings.· Entity summary generated by Graphiti

RELATIONS
43Relations
CURRENT
42Current
EVENTS
1Events
REPORTS
4Source reports
KNOWLEDGE ATLAS0 entities · 0 relations
No public neighbors yet
Explore all published links of a node; previously discovered relations remain visible.

Relation timeline

BY VALID-FROM DATE
  1. Mistral Large 4TRAILSChina

    Mistral Large 4 trails China overall

    Current1 sourceChina
  2. Mistral Large 4HAS SCORE 51 PERCENT ONArtificial Analysis Cyber Index

    Mistral Large 4 scores 51% on the Artificial Analysis' Cyber Index

    Current1 sourceArtificial Analysis Cyber Index
  3. Mistral Large 4OUTPERFORMS AT CODINGGLM-5.3

    Mistral Large 4 outperforms GLM-5.3 at coding

    Current1 sourceGLM-5.3
  4. Mistral Large 4OUTPERFORMS AT CODINGDeepseek V4.1-Flash

    Mistral Large 4 outperforms DeepSeek V4.1 Flash at coding

    Current1 sourceDeepseek V4.1-Flash
  5. Mistral StudioTESTS SPECIAL VERSIONMistral Large 4

    Orange Cyberdefense is testing a special version of Mistral Large 4 for penetration testing

    Current1 sourceMistral Studio
  6. Mistral Large 4REDUCES TOKENS BY PERCENTAGEGLM-5.3

    Mistral Large 4 uses 47 percent fewer tokens per task than GLM‑5.3

    Current1 sourceGLM-5.3
  7. Mistral Large 4REFUSES ONStrongREJECT

    Mistral Large 4 has a higher refusal rate on StrongREJECT than other open‑weight models

    Current1 sourceStrongREJECT
  8. Mistral Large 4OUTRANKSArtificial Analysis' Coding Agent Index

    Mistral Large 4 outranks all other open models for agentic coding tasks on Artificial Analysis’s Coding Agent Index

    Current1 sourceArtificial Analysis' Coding Agent Index
  9. Mistral Large 4AVAILABLE IN VARIANTGGUF

    Mistral Large 4 also has GGUF variants

    Current1 sourceGGUF
  10. Mistral Large 4DISTRIBUTED UNDERMistral Ecosystem Open-Weight License

    Mistral Large 4 is distributed under the Mistral Ecosystem Open‑Weight License

    Current1 sourceMistral Ecosystem Open-Weight License
  11. Mistral Large 4COMPARES PERFORMANCE WITHGLM-5.3

    Mistral Large 4’s performance is comparable to the Chinese open‑weight competitor GLM‑5.3

    Ended 2026-10-291 sourceGLM-5.3
  12. Mistral Large 4IS AKALe Chonk

    Mistral Large 4 is also known as Le Chonk

    Current3 sourcesLe Chonk
  13. Mistral Large 4USESML4

    Mistral Large 4 uses the Pixtral vision encoder

    Current1 sourceML4
  14. Hugging FacePROVIDES WEIGHTS FORMistral Large 4

    Hugging Face provides the weights for Mistral Large 4 (ML4)

    Current1 sourceHugging Face
  15. Mistral Large 4USESMistral Voxtral-Mini-4B-Realtime

    Mistral Large 4 uses LeRope for rotated position embeddings

    Current1 sourceMistral Voxtral-Mini-4B-Realtime
  16. Mistral Large 4FROMMistral

    Mistral Large 4 is from Mistral

    Current1 sourceMistral
  17. Mistral Large 4AVAILABLE IN VARIANTFP8 NVFP4

    Mistral Large 4 is released in the FP8 NVFP4 variant

    Current1 sourceFP8 NVFP4
  18. MistralRELEASES WEIGHTS OFMistral Large 4

    Mistral releases the weights for its trillion‑parameter open source AI model Mistral Large 4 (ML4)

    Current1 sourceMistral
  19. Mistral Large 4REFUSES ONAgentHarm

    Mistral Large 4 has a higher refusal rate on AgentHarm than other open‑weight models

    Current1 sourceAgentHarm
  20. Mistral Large 4USESMistral Studio

    Mistral Large 4 can be fine‑tuned with TRL

    Current1 sourceMistral Studio
  21. Mistral Large 4REFUSES ONJailbreakBench

    Mistral Large 4 has a higher refusal rate on JailbreakBench than other open‑weight models

    Current1 sourceJailbreakBench
  22. Mistral Large 4RANKED INMLX

    Mistral Large 4 ranks in Dense 200 visual grounding

    Current1 sourceMLX
  23. Mistral Large 4OUTPERFORMSVLMEvalKit

    Mistral Large 4 outperforms Qwen3.8 Max on Coding Agent Index

    Current1 sourceVLMEvalKit
  24. Mistral Large 4OUTPERFORMSGPT-6 Astra

    Mistral Large 4 outperforms GPT-6-Astra on Dense 200 visual grounding

    Current1 sourceGPT-6 Astra
  25. Mistral Large 4BUILT WITHMistral Forge

    Mistral Large 4 was built using the same training, customization, and RL environment as Mistral Forge

    Current1 sourceMistral Forge
  26. Mistral Large 4TRAINED ONNVIDIA Grace Blackwell GPUs

    Mistral Large 4 was trained on NVIDIA Grace Blackwell GPUs

    Current1 sourceNVIDIA Grace Blackwell GPUs
  27. Mistral Large 4RANKED INFalcon-ASR

    Mistral Large 4 ranks in the Artificial Analysis Cyber Index

    Current1 sourceFalcon-ASR
  28. Mistral Large 4ACHIEVES SCORE ONGPT‑5.4

    Mistral Large 4 achieves a score on DeepSWE v1.1

    Current1 sourceGPT‑5.4
  29. Mistral Large 4HAS ALTERNATIVE NAMEle Chonk

    Mistral Large 4 has the alternative name le Chonk

    Current1 sourcele Chonk
  30. Mistral Large 4PERFORMSML Intern

    Mistral Large 4 can perform a Hartree–Fock simulation in one shot

    Current1 sourceML Intern
  31. Mistral Large 4AVAILABLE ONMistral Studio

    Mistral Large 4 is available on Mistral Studio

    Current1 sourceMistral Studio
  32. Mistral Large 4ACHIEVES SCORE ONMila Quacquarelli

    Mistral Large 4 achieves a score on Terminal-Bench 4

    Current1 sourceMila Quacquarelli
  33. Mistral Large 4ACHIEVES SCORE ONMAPL-EMIT

    Mistral Large 4 achieves a score on SWE-Atlas-QnA

    Current1 sourceMAPL-EMIT
  34. Mistral Large 4TRAINED INPharos-CY

    Mistral Large 4 was trained in Mistral’s own datacenters

    Current1 sourcePharos-CY
  35. Mistral Large 4RANKED INSerdar Ozkan

    Mistral Large 4 ranks in AA-Briefcase

    Current1 sourceSerdar Ozkan
  36. Mistral Large 4RANKED INDestry Damayanti

    Mistral Large 4 ranks in AutomationBench

    Current1 sourceDestry Damayanti
  37. Mistral Large 4RANKED INML Intern

    Mistral Large 4 ranks in SciCode-Verified for scientific coding

    Current1 sourceML Intern
  38. Mistral Large 4OUTPERFORMSProtNLM

    Mistral Large 4 outperforms DeepSeek V4 Pro 0813 on Coding Agent Index

    Current1 sourceProtNLM
  39. Mistral Large 4HAS SCORE 19 7 PERCENT ONMistral OCR

    Mistral Large 4 scores 19.7% on the FrontierCode 2.0 benchmark

    Current1 sourceMistral OCR
  40. Mistral Large 4HAS SCORE 41 PERCENT ONTerminal-Bench 4

    Mistral Large 4 has a score of 41% on Terminal‑Bench 4.0

    Current1 sourceTerminal-Bench 4

Relations are facts Graphiti extracted from reports, not verbatim quotes. An end date means later reporting showed the relation no longer holds.

Events 1

Source reports 4