Mistral Small 4

How it is built

Mistral Small 4 has 119B parameters, and Mistral gives two counts for the active part. The documentation and the model card say 6.5B per token, about 5.5%.¹¹1Mistral AI, Mistral Small 4 119B A6B model card, Hugging Face. https://huggingface.co/mistralai/Mistral-Small-4-119B-2603. Arithmetic: 6.5 / 119 = 5.5%. ²²2Mistral AI documentation, Mistral Small 4 model card, version 26.03, released 16 March 2026, prices per million tokens, read 29 September 2026. https://docs.mistral.ai/models/model-cards/mistral-small-4-0-26-03 The launch post says 6B, or 8B with the embedding and output layers.³³3Mistral AI, launch post for Mistral Small 4, 16 March 2026. https://mistral.ai/news/mistral-small-4/

All 36 layers are mixture of experts layers, with no dense layer in front.⁴⁴4Mistral AI, params.json in the official repository: n_layers 36, first_k_dense_replace 0, num_experts 128, num_shared_experts 1, num_experts_per_tok 4, kv_lora_rank 256, q_lora_rank 1024, max_position_embeddings 1048576, qformat_weight fp8_e4m3, vision encoder with 24 layers. https://huggingface.co/mistralai/Mistral-Small-4-119B-2603/blob/main/params.json The expert counts match Mistral Large 3: 128 routed experts and 1 shared expert per layer, and 4 routed experts for each token.⁴⁴

Keys and values are stored in a compressed latent of rank 256.⁴⁴ This is multi-head latent attention, and the vLLM command on the card selects an MLA backend for it.¹¹

Images go through a vision encoder of 24 layers.⁴⁴

What is new

The reasoning models Mistral used to call Magistral now live in the same weights as Instruct. So does Devstral.¹¹

Reasoning is chosen per request with reasoning_effort. At none the replies read like Mistral Small 3.2. At high they run as long as the old Magistral models.¹¹

Against Mistral Small 3, Mistral reports a 40% cut in end-to-end completion time in a latency-optimized setup.¹¹ Optimized for throughput instead, it serves 3 times as many requests per second.¹¹

Mistral also counts output length. On AA LCR, Small 4 with reasoning scores 0.72 with 1.6K characters of output, and Mistral says Qwen models need 3.5 to 4 times more for comparable scores.¹¹ On LiveCodeBench it beats GPT-OSS 120B while writing 20% less.¹¹

Running it

The weights are FP8, under Apache 2.0.¹¹⁴⁴ That comes to about 119 GB (119B x 1 byte).⁵⁵5Arithmetic: 119B x 1 byte = 119 GB, weights only, before the KV cache. https://huggingface.co/mistralai/Mistral-Small-4-119B-2603/blob/main/params.json The card's vLLM command spreads them over 2 GPUs.¹¹ An NVFP4 checkpoint and an EAGLE head for speculative decoding are also published.¹¹

The config allows 1,048,576 positions.⁴⁴ Both the documentation and the card stop at 256k tokens.¹¹²²

mistral-small-2603 costs $0.15 per million input tokens and $0.6 per million output tokens on the API.²²

Six weeks later came Mistral Medium 3.5, a dense model of 128B.⁶⁶6Mistral AI documentation, Mistral Medium 3.5 model card, released 28 April 2026. https://docs.mistral.ai/models/model-cards/mistral-medium-3-5-26-04. Arithmetic: 16 March to 28 April 2026 is 43 days, about six weeks. It is only a little larger, and it uses every parameter on every token.