Mistral Pixtral Large 2.5 reaches SOTA open multimodal reasoning on MMMU-Pro and MathVista-Hard
[AI-NEWS]
Date: 2026-02-25
Content : Pixtral Large 2.5 (open-weight) sets new open-model records on MMMU-Pro (visual reasoning) and MathVista-Hard (math + vision), surpassing previous open SOTA by 9–14 points while maintaining efficient inference on 8×A100 or consumer multi-GPU setups.
Keywords : multimodal reasoning, MMMU-Pro, MathVista, Pixtral Large 2.5, open-weight vision-language model
Next article
No more
new
Mistral Pixtral Large 2.5 reaches SOTA open multimodal reasoning on MMMU-Pro and MathVista-Hard
Anthropic Claude 5 Sonnet demonstrates 14-day autonomous scientific literature review + hypothesis generation
xAI Grok-5.1 integrates real-time multi-agent debate for improved truthfulness and calibration
Google DeepMind AlphaFold 4 predicts all known protein complexes with near-experimental accuracy
OpenAI o4-proto-4 achieves 82% autonomous resolution on full SWE-Bench Verified + Multi-File
Mistral Mathstral 2.0 sets new open-source SOTA on MATH and GSM8K-Hard
Anthropic Claude 4.4 Opus demonstrates 7-day autonomous research agent on novel scientific hypothesis
xAI Grok-5 preview adds native causal video generation with physics consistency
OpenAI o4-proto-3 achieves closed-loop self-correction on 100+ cycle software debugging tasks
Mistral Devstral: Specialized 24B Model Optimized for Developer Tool-Chain Integration