Pixtral 12B

Pixtral 12B (checkpoint 2409) was Mistral AI's first multimodal model, launched September 2024: a 12B-parameter multimodal decoder plus a 400M-parameter vision encoder, built on the Mistral Nemo 12B backbone. Unlike adapter-bolted VLMs, it is natively multimodal — trained on interleaved image and text data — with variable image-size support and a 128K-token sequence length.

Benchmarks are provider-reported (all models re-evaluated through Mistral's own pipeline, per the model card): MMMU 52.5, MathVista 58.0, ChartQA 81.8, DocVQA 90.7, VQAv2 78.6, plus instruction-following scores (MM MT-Bench 6.05, Text MT-Bench 7.68). Notably the card's own table shows DocVQA trailing Qwen2-VL 7B (90.7 vs 94.5) — honest framing.

Released under Apache 2.0 (verified in the HF model card and confirmed by a Mistral spokesperson), so it is freely usable, fine-tunable, and commercially deployable — one of the most permissively licensed early open VLMs. Available via Mistral's la Plateforme API and Le Chat, plus self-host. Note: Mistral's current pricing page (checked 2026-10-13) no longer shows an explicit pixtral-12b price row — confirm current availability/pricing before quoting; launch-era pricing was widely reported as $0.15/1M input + $0.15/1M output (provider-claimed).

Provider-reported (Mistral's own eval pipeline): MMMU 52.5, MathVista 58.0, ChartQA 81.8, DocVQA 90.7, VQAv2 78.6; MM MT-Bench 6.05, Text MT-Bench 7.68, MM IF-Eval 52.7, Text IF-Eval 61.3.

Independent: Launch press confirmed the Apache 2.0 license and community reception; no large-scale independent replication of the exact benchmark table was found.