Ministral 3 3B 2512

Aireleasetracker records Ministral 3 3B-25-12 as an open-weight release from Mistral on December 2, 2025, arriving 165 days after Mistral Small 3.2. It confirms input and output pricing of $0.10 per million tokens across listed Mistral endpoints, with an EU variant priced at $0.11 per million tokens. Context length on this page is reported as 128K tokens per endpoint. Three Mistral serving tiers are listed: standard, ZDR (zero data retention), and an EU regional endpoint, all with 128K context. Pricing was verified against mistral.ai on August 18, 2026, with provider rates fetched from openrouter.ai on October 15, 2026. The page frames the weights as downloadable for local self-hosting given its open-weight status.

OpenRouter's catalog describes Ministral 3 3B 2512 as the smallest model in the Ministral 3 family, a tiny language model with vision capabilities released December 2, 2025. Modalities are text and image in, text out, with a 131K context window and listed pricing of $0.10 per million tokens for both input and output. Weighted-average effective prices paid are $0.06845 input and $0.09526 output per million tokens. Live performance figures show throughput of 173 tok/s and P50 latency of 0.31s on the best Mistral endpoint, with end-to-end latency averaging 0.74s. Tool call error rate on the Mistral EU endpoint is 0.00 percent, and 30-day uptime reaches 99.97 to 100.00 percent across listed endpoints. Standard routing options include Balanced, Nitro, Floor, and Exacto modes for price, speed, or tool-calling accuracy.

Modelgrep describes Ministral 3 3B 2512 as a 3.8B-parameter model under Apache-2.0, with a 131K context window and up to 105K max output tokens. It supports text and image input modalities producing text output, with prompt caching at $0.010 per million tokens and no moderation. The page lists three Mistral providers at $0.100 input and $0.100 output per million tokens with 131K context and full uptime. Speed measurements on Modelgrep report approximately 201 tokens per second generation throughput and 315ms time-to-first-token at p50, ranking it the 34th fastest tracked model. Design Arena Elo scores include 1029 for Website, 964 for 3D, and 953 for Game Dev across 2,319 tournaments. The open weights path is mistralai/Ministral-3-3B-Instruct-2512 on Hugging Face.

Magica's model profile confirms Ministral 3 3B 2512 was developed by Mistral and released December 2, 2025. Input context window is listed at 131.1K tokens with an output token limit of 104.9K tokens, and pricing is $0.10 per million tokens for both input and output with $0.01 per million cache reads. The page lists function calling and structured output among supported key features, with text and image inputs producing text outputs. Composite benchmark indices reported by Magica show an Intelligence Index of 4.8 and Coding Index of 4.8 from Artificial Analysis, plus an Agentic Index of just 0.8 indicating weaker multi-step task performance. The Design Arena best score is 1,029 Elo overall, with the Website subcategory at 122 (38.2 percent win rate).

Benchleader tracks Ministral 3 3B released December 2, 2025 with an index of 37.0 and overall rank 723. Detailed benchmark scores include GPQA Diamond at 35.8 percent, Humanity's Last Exam at 5.4 percent, SciCode at 15.3 percent, and MMMU-Pro multimodal score of 38.1 percent. The AA Intelligence Index v4.3.2 composite lands at 4.8, rank 598 among all tracked configurations. Agentic and tool-use evaluation results show τ²-Bench Telecom at 24.9 percent, τ³-Banking at 4.7 percent, Terminal-Bench Hard at 0.0 percent, and GDPval-AA v2.1 at 0.0 percent. Knowledge benchmarks include AA-Omniscience accuracy of 9.0 percent and non-hallucination of 19.8 percent, while instruction-following IFBench reaches 26.8 percent and AA-LCR long-context benchmark scores 17.0 percent.

Mistral reached general availability for Ministral 3 3B (version v25.12) on December 2, 2025, releasing it under Apache 2.0 as the smallest, most efficient entry in the Ministral 3 family. The official Mistral docs page lists a 256K context window with $0.10 per million tokens for both input and output in USD pricing. Document QnA, Function Calling, Structured Outputs, chat completions and batching are exposed across /v1/chat/completions, /v1/conversations and /v1/batch endpoints. The model is positioned for edge deployment and described as offering robust language and vision capabilities in a compact package. It supports a plus-one variant labeled ministral-3b-2512 alongside the base model on the same documentation page. Mistral also cross-references Z.ai GLM 5.3 and 5.2 plus Shieldstral 1.0 as related catalog entries on the same page.