AI Potluck
Model components / Fine-tuned / chat models

Mistral-7B-Instruct-v0.2

Mistral AI

Instruction-tuned 7B dense chat model from Mistral with a 32K context, fine-tuned from the Mistral 7B base. It was the first widely adopted European open instruct release, in December 2023, and remains a common 7B chat baseline in evaluation papers and fine-tuning tutorials.

Superseded by later point releases and by the Ministral and Mistral Small lines. The card carries no benchmark table and links only the base-model paper, so the capability figures come from the paper and the La Plateforme launch post. Verified 2026-08-13 via the model card.

Openness

3 high confidence
3.0
weights
open(Apache-2.0, on HF)
data
closed
code
closed(no training/post-training pipeline)
license
Apache-2.0(OSI permissive)

Apache-2.0 open weights, with the instruction-tuning data and recipe undisclosed, so this is open weights rather than open source.

  • https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2 recorded 2026-08-13

    `license:apache-2.0` in the repo tags and `"gated":false` in the embedded repo state, so the 7B weights download without a barrier; the card describes the instruction fine-tune in prose and publishes neither the mixture nor any training or post-training code

Adoption

4 high confidence
4.0

The single declared SKU reads 1,206,680 Hugging Face downloads in the trailing 30 days, inside the 1M-10M band, and it carries a very large derivative ecosystem - roughly 1,100 finetunes, 1,200 adapters and 351 merges. Still a heavily used permissive 7B instruct base years after release.

Capability

2 medium confidence
2.0

Excellent for a 7B in 2023-24; clearly below 2026 frontier and below modern small instruct models (Phi-4-mini, Qwen 3.x small). Capability low on a frontier-anchored scale. The MT-Bench 7.6 comes from Mistral's La Plateforme post and is specific to v0.2 - the paper's Table 3 gives v0.1-Instruct 6.84 - and the MMLU 60.1% belongs to the Mistral 7B BASE model in Table 2 rather than to this instruct tune. The model card itself carries no benchmark table.

  • https://mistral.ai/news/la-plateforme recorded 2026-08-14

    Mistral's La Plateforme launch post - "Mistral-tiny. Our most cost-effective endpoint currently serves Mistral 7B Instruct v0.2, a new minor release of Mistral 7B Instruct. Mistral-tiny only works in English. It obtains 7.6 on MT-Bench."

  • https://arxiv.org/pdf/2310.06825 recorded 2026-08-14

    Mistral 7B paper full text. Table 2 gives the BASE model at MMLU 60.1%, HellaSwag 81.3%, WinoGrande 75.3%; Table 3 gives "Mistral 7B Instruct" (v0.1) an MT-Bench of 6.84 +/- 0.07. Neither table covers the v0.2 instruct tune.

  • https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2 recorded 2026-08-13

    7B instruct model card, 32k context; links arxiv:2310.06825 "Mistral 7B", the base-model paper. No benchmark table on the card.

Verified 2026-08-13