Mixtral 8x22B

Claims and evidence

• Released 17 April 2024Other sources give: 10 April 2024 [3] [4]Warning The base weights were first shared through a magnet link that Mistral AI posted on X (post timestamp 2024-04-10T01:20:38Z, still 9 April in American time zones); the Hugging Face base card says its weights are based on that torrent release. The announcement, the instruct model and API access followed on 2024-04-17. ● Primary source[1]page date [2]April 17, 2024 [5]page header [6]Model Lifecycle: Mixtral 8x22B Base, Mixtral 8x22B Instruct
• Deprecated 30 November 2024 ● Primary source[7]Deprecated table, Mixtral 8x22B 0.1-0.3 (open-mixtral-8x22b)
• Retired 30 March 2025Retirement on Mistral's platform; the open weights remain downloadable. ● Primary source[7]Deprecated table, Mixtral 8x22B 0.1-0.3 (open-mixtral-8x22b) [5]Retirement date [6]Model Lifecycle: Mixtral 8x22B Base, Mixtral 8x22B Instruct

• Status AvailableThe Apache 2.0 base and instruct weights are still downloadable from Mistral AI's Hugging Face organisation. On Mistral's own platform the model is retired (models overview, Deprecated table; legal centre). ● Primary source[4] [8] [5]Weights table
• Replaced by Mistral Small 4 ● Primary source[7]Deprecated table, Mixtral 8x22B 0.1-0.3, Alternative [5]Replacement

• Successor of Mixtral 8x7B EditorialEditorial link along the Mixtral line: the announcement presents Mixtral 8x22B as Mistral's latest open model and a natural continuation of its open model family, and charts it next to Mixtral 8x7B, but names no predecessor. ● Primary source[1]Efficiency at its finest; Figure 1

• Change · Size Larger sparse mixture-of-experts model: 141B total and 39B active parameters, against 46.7B and 12.9B for Mixtral 8x7B.Compared with Mixtral 8x7B ● Primary source[1]first paragraph [9]Pushing the frontier of open models with sparse architectures
• Change · Context length Context window of 64K tokens, up from 32k for Mixtral 8x7B.Compared with Mixtral 8x7B ● Primary source[1]strengths list [9]capabilities list
• Change · Tool use Native function calling, which Mistral AI lists among the model's strengths.Compared with Mixtral 8x7B ● Primary source[1]strengths list [8]Function calling example

• Input text ● Primary source[5]Modalities
• Output text ● Primary source[5]Modalities
• Feature Function calling ● Primary source[1]strengths list [8]Function calling example
• Feature MultilingualEnglish, French, Italian, German and Spanish. ● Primary source[1]strengths list
• Open weights YesApache 2.0. ● Primary source[1]Truly open [4] [8]
• Context window 64K tokens ● Primary source[1]strengths list [5]Context; Weights table
• Parameters 141B total, 39B activeSparse mixture of experts. ● Primary source[1]first paragraph [5]Weights table
• Access Open-weights download, API ● Primary source[1]Truly open; last paragraph [2]April 17, 2024

Lineage

Variants

No variants recorded in this record.

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from Mistral AI on AI Radar →