Mistral Small 4 119B · Atomic Dynamic GGUF

Atomic Dynamic (AD) quants of Mistral AI's Mistral Small 4 (119B parameters, 6.5B active), built from Mistral's original weights for the 64 GB and 96 GB memory classes.

• You can now run Mistral Small 4 in Atomic Chat, with reasoning off by default and one switch to turn it on.
• Every file here was measured against the original weights on one stand, next to Unsloth's file of the same size.

---

Pick a file

KL divergence is how far the quantized model's predictions drift from the original weights: lower is better, zero means identical. top-1 is how often it picks the same next token the original would have picked. Raw logs are in the metrics repo.

AD- marks an Atomic Dynamic layout; the rest of the name is the type of the routed experts' gate and up projections, the largest tensor group. Compare files by size on disk, not by name. The vision projector is shared by every file: mmproj-Mistral-Small-4-119B-2603-F16.gguf (0.86 GB) or -Q8_0 (0.47 GB).

Against Unsloth at the same size

Each row compares two files of about the same size, measured with the same reference logits, text and build. "Ours vs theirs" is the paired difference in mean KL divergence over the 86 chunks of the held-out text, with a bootstrap interval over chunks.

We publish only files that beat the Unsloth file of their size. At the 128 GB size (Unsloth's UD-Q5_K_XL, 89.45 GB) our best try read 2.0% lower (95% CI -4.7 to +1.2), which is noise, so there is no file of that size here. Even plain Q8_0 (126 GB) reads KL divergence 0.041 and top-1 92.3% on this model: above about 5 bits per weight the files differ little.

What is different in these files

Every file keeps attention, the shared expert and the router side at q8_0 and changes only the routed experts, as Unsloth's files do. What differs:

• The first blocks. With 128 experts and 4 active, a small error early changes which experts every later block picks. Moving bytes to the blocks with the most activation energy (the last ones, by our importance matrix) changed nothing against Unsloth's choice of blocks (4, 28, 32, 35). Moving them to the first four did.

• The importance matrix. Ours is 3.8M tokens through this model's own chat template (Unsloth's files carry one of 102 chunks, by their metadata). On Unsloth's own tensor types it alone gave 4.2% lower KL divergence at the 50 GB size (95% CI -8.0 to -1.1) and 2.2% at 69 GB (within noise).

• AD-IQ3_XXS (49.64 GB): Unsloth's tensor types for this size, block for block (gate/up iq3_xxs, down iq3_s with iq4_xs on blocks 4, 28, 32 and 35, token_embd and output q6_k), with our importance matrix. At this size none of our six other layouts beat it: moving the iq4_xs blocks to the last blocks or to the first ones, raising the first blocks' gate/up, or the head at q8_0 read the same or worse. 4.2% lower KL divergence than UD-Q3_K_S, lower in 58 of 86 chunks.

• AD-IQ4_XS (69.22 GB): gate/up iq4_xs instead of q4_k, which frees 2.4 GB; blocks 0-3 get gate/up q5_k and down q6_k, blocks 4-6 down q6_k, the rest down q4_k; token_embd q8_0, output q6_k. 0.25 GB smaller than UD-Q4_K_S, 7.8% lower KL divergence, lower in 72 of 86 chunks.

Running it

llama-server -m Mistral-Small-4-119B-2603-<FILE>-00001-of-00002.gguf \
  --mmproj mmproj-Mistral-Small-4-119B-2603-F16.gguf --jinja -ngl 99 -c 32768 --temp 0.1

Reasoning is off unless a request asks for it (the template takes none or high):

curl http://localhost:8080/v1/chat/completions -d '{"messages": [{"role": "user", "content": "..."}],
  "chat_template_kwargs": {"reasoning_effort": "high"}, "temperature": 0.7}'

Temperature 0.1 without reasoning and 0.7 with it are Mistral's recommended settings. The chat template inside the files is Mistral's, with one change: the default system prompt's {today} is filled with the current date.

Every published file passed the same behaviour check in llama-server (this build, --jinja, our projector): answers in English and German with reasoning off, the reasoning switch (high: a thinking block, then the right answer), a tool call with the right function and arguments, the colour of a shape in an image, and a number found in the middle of a 32K-token prompt.

How we measured

Reference: Mistral's FP8 checkpoint upcast to BF16 and run unquantized. We checked Unsloth's BF16 GGUF against the FP8 checkpoint tensor by tensor (15 of 15 sampled tensors bit-identical) and use it as the reference and as the input of our quants. Held-out text: eval_neutral from our calibration dataset (30 languages, 86 chunks of 4096 tokens), never used for calibration. Every file in the comparison, ours and Unsloth's, ran with the same reference logits, text, context and llama.cpp build (0893ec98) on one 8x A100-SXM4-40GB box. Differences are paired by chunk with a bootstrap over the 87 chunks.

Our importance matrix: 3.8M tokens rendered through this model's own chat template (agentic, code, reasoning with [THINK] blocks, 30+ languages, structured text and a sweep of the whole vocabulary) at context 512, collected on the BF16 weights. (A second pass over long documents at context 32768 crashed in llama-imatrix and is not in it.) It is imatrix/imatrix.gguf in the metrics repo; the corpus is builds/mistral-small-4-moe in the calibration dataset.

Model details

Mistral Small 4 (mistral4 in llama.cpp): 36 layers, hidden size 4096, multi-head latent attention, 128 routed experts (4 active) plus one shared expert in every layer, vocabulary 131,072, untied output head, a Pixtral-style vision encoder. Mistral gives 256K tokens of context; we measured at 4096. Support landed in llama.cpp on 16 March 2026 (PR #20649).

License

Mistral Small 4 is released by Mistral AI under the Apache License, Version 2.0 (copy in LICENSE). Base model: mistralai/Mistral-Small-4-119B-2603.

Quantized by AtomicChat.