Magistral Small 1.2
Sep 2025
Retired 2026-07-31Replace withMistral Small 4Newer version
MiMo v2.6 ProCheaper, as good
GPT-6 LunaBetter
Qwen3.8 27BOpen weights
Mistral Small 4Newer versionMiMo v2.6 ProCheaper, as goodGPT-6 LunaBetterQwen3.8 27BOpen weightsMagistral Small 1.2Compare inon
Benchmarks of Magistral Small 1.221
BenchmarkByScoreRank
Overall1
Knowledge4
Science2
Coding1
Agents3
Multimodal1
Health & law1
Documents1
Legacy · saturated or no longer reported7
No benchmark matches
Rankingin thinking mode?
Retired models are left out of the current rankings.
Alternatives?
ModelScoreCostPer taskWhat changes
CheaperMiMo v2.6 ProthinkingOpen×5.4-28%$0.13Memory (4-bit): 14 GB → 615 GB+4BetterGPT-6 Lunalow×2.5-73%$0.005Function calling only with none · Chat Completions+3OpenQwen3.8 27Bxhigh×3.9+27%$1.01Memory (4-bit): 14 GB → 17 GB+5
CheaperDeepSeek V4.1 Flash 1017maxOpen×5.2-30%$0.26Memory (4-bit): 14 GB → 458 GB+5
CheaperClaude Haiku 5.5max×5-73%$0.21Remove sampling parameters+5ModelScoreCostPer taskWhat changes
MiMo v2.6 ProthinkingOpen×5.4-28%$0.13Memory (4-bit): 14 GB → 615 GB+4DeepSeek V4.1 Flash 1017maxOpen×5.2-30%$0.26Memory (4-bit): 14 GB → 458 GB+5Claude Haiku 5.5max×5-73%$0.21Remove sampling parameters+5
MiniMax M3.1 Flashthinking×5-65%$0.22Complete model response must be appended
GLM 5.3 FlashmaxOpen×4.9-68%$0.28Memory (4-bit): 14 GB → 193 GB+5ModelScoreCostPer taskWhat changes
GPT-6 Lunalow×2.5-73%$0.005Function calling only with none · Chat Completions+3GPT-5.6 Lunalow×2.4-40%$0.01Mistral Small 4thinkingOpen+31%-65%$0.015Memory (4-bit): 14 GB → 72 GBMiMo v2.5thinkingOpen×2.5-77%$0.019Memory (4-bit): 14 GB → 186 GB
Granite 4.2 8BthinkingOpen+29%-86%$0.024No image inputModelScoreCostPer taskWhat changes
Qwen3.8 27BxhighOpen×3.9+27%$1.01Memory (4-bit): 14 GB → 17 GB+5IK2 Horizon MoVA 36B A4BthinkingOpen×2.9+133%$1 / $4Memory (4-bit): 14 GB → 22 GB+1Qwen 3.5 27BthinkingOpen×2.7+10%$0.30 / $2.40Memory (4-bit): 14 GB → 17 GBAG9v3 39A5BOpen×2.5——Memory (4-bit): 14 GB → 23 GB+1Qwen3.6 27BthinkingOpen×2.5+80%$0.62Memory (4-bit): 14 GB → 17 GBModelScoreCostPer taskWhat changes
Mistral Large 4thinking×4.5—$1.13Infinite whitespace stream+4Mistral Medium 3.5highOpen+65%—$0.50Memory (4-bit): 14 GB → 77 GBMistral Small 4thinkingOpen+31%—$0.015Memory (4-bit): 14 GB → 72 GBMistral Large 3non-thinkingOpensame—$0.031Memory (4-bit): not verifiedMagistral Medium 1thinkingsame——Context: 128K → 40K+1Devstral Mediumnon-thinkingOpensame—$0.40 / $2Memory (4-bit): not verified+1Magistral Small 1thinkingOpensame—$0.70 / $1.16No image inputMistral Small 3non-thinkingOpen-22%—$0.05 / $0.08Context: 128K → 32K+1All Mistral models →Run it yourself
estimateFits on✗ 8 GB (phone, laptop)✓ 16 GB · compressed✓ 24 GB GPU · compressed✓ 48 GB GPU · 8-bit✓ 80 GB GPU · full quality
LicenceApache-2.0free for commercial use
History
2 versionsBenchmarks
all categoriesChat & Knowledge
AA Omniscience Index-65.1
GPQA Diamond66.3
Coding
Agents & Tool Use
τ²-bench · AA27.8
Writing & Language
Vision & Documents
FAQ
How much does Magistral Small 1.2 cost?›
$0.50 per 1M input tokens and $1.50 per 1M output tokens (cheapest host: AWS Bedrock).
Is Magistral Small 1.2 free to use?›
There is no free API; the cheapest paid price is $0.50 per 1M input tokens. The weights are open, so you can run it yourself.
How can I use Magistral Small 1.2?›
Through the API of AWS Bedrock, Featherless.
What is the context window of Magistral Small 1.2?›
128K tokens.
When was Magistral Small 1.2 released? Is there a newer version?›
Released 2025-09-17. It has been superseded by Mistral Small 4.
Is Magistral Small 1.2 open source?›
Yes, the weights are open (Apache-2.0 licence).
Can I run Magistral Small 1.2 locally?›
The weights are open and can be self-hosted; hardware needs are not confirmed yet.
Can Magistral Small 1.2 read images?›
It accepts text and images as input.
What is Magistral Small 1.2?›
Magistral Small 1.2 is an open-weights language model from Mistral, released Sep 2025, the successor to Magistral Small 1.
What are the best alternatives to Magistral Small 1.2?›
MiMo v2.6 Pro (cheaper), GPT-6 Luna (better), Qwen3.8 27B (open weights).