Ministral 3 8B

Artificial Analysis rates Ministral 3 8B at 5 on the Intelligence Index, marking it below average among comparable open-weights non-reasoning peers. Output throughput is reported at roughly 99.9 tokens per second, faster than the class median, while verbosity is flagged as somewhat high at around 20M tokens generated across the intelligence evaluation. Pricing sits at $0.15 per 1M input and $0.15 per 1M output tokens, with a 90% cache discount. The model is described as supporting text and image inputs with text outputs, a 256k-token context window, 8B total parameters, and an Apache 2.0 license, with weights available on Hugging Face. The page notes a possible separate reasoning variant while explicitly profiling this snapshot as the non-reasoning configuration, and compares the model within the small open-weights peer class.

Benchmark Atlas catalogs Ministral 3 8B as an 8.9B-parameter dense multimodal Mistral model exposed via the API as ministral-8b-2512 and linked to the Instruct weights, with separate Base and Reasoning checkpoints published. Its configuration reaches 262,144 positions through YaRN scaling from an original 16,384-position window. Artificial Analysis scores include an Intelligence Index of 5.5 (v4.3 and v4.3.2), AA-Coding Index 9.7, MMLU-Pro 64%, GPQA Diamond 47.1, and LiveCodeBench 30.3. Aggregated results span roughly forty evaluations with headlines attributed largely to Artificial Analysis, alongside independent scores such as FormationEval 86.9 and Global MMLU Lite 66.3. The page also surfaces snapshot leaderboard placements like FormationEval at 45/60 and GDPval-AA v2 (normalized) at 0.0, giving a multi-metric view of the model's standings.

Upend.ai's listing confirms Ministral 3 8B 2512 as a Mistral AI multimodal model released on 2025-12-02, accepting text and image inputs and producing text outputs. It documents a 262,144-token context length and supports parameters including frequency/presence penalty, max tokens, response format, seed, stop, structured outputs, temperature, tool choice, tools, and top_p. Defaults are temperature 0.3 and top_p 0.95. The listing ties the model ID mistralai/ministral-8b-2512 to the Ministral 3 family and outlines sample use cases spanning language understanding, multimodal processing, structured outputs, and parameter-tuned responses. It is presented as part of a broader Mistral model catalog view.

Mistral AI released Ministral 3 8B (API ID ministral-8b-2512, version v25.12) to general availability on December 2, 2025 under an Apache 2.0 license. The official Mistral docs page frames the model as a compact, edge-optimized member of the Ministral 3 family built for local and diverse hardware setups. Multimodal text and vision capability is highlighted alongside a 256k token context window. Hosted Mistral pricing is set at $0.15 per million tokens for both input and output. Supported endpoints include /v1/chat/completions, /v1/conversations, structured outputs, function calling, document QnA, prefix caching, and /v1/batch batching. The page also lists related Mistral offerings such as Shieldstral 1.0 and Z.ai GLM 5.x variants for comparison.