unsloth
Mistral Small 4 119B 2603
Mistral Small 4 119B 2603 is a 119.0B mixture-of-experts mistral4 model from unsloth, with about 21.6B parameters active per token. Configured context length is 1,048,576 tokens. License: apache-2.0. Quant sizes are the GGUF file bytes published on Hugging Face when a file exists.
Parameters119B / 21.6B active
Context1024K
Attentionmla
Licenseapache-2.0
36 layers · 32 KV heads · head dim 128 · GGUF repo · config
Quant file sizes
| Quant | File size | Source |
|---|---|---|
| BF16 | 221.6 GB | Hugging Face file bytes |
| UD-Q8_K_XL | 123.5 GB | Hugging Face file bytes |
| Q8_0 | 117.8 GB | Hugging Face file bytes |
| UD-Q6_K_XL | 100.8 GB | Hugging Face file bytes |
| UD-Q6_K | 92.6 GB | Hugging Face file bytes |
| UD-Q5_K_XL | 83.3 GB | Hugging Face file bytes |
| UD-Q5_K_M | 83.0 GB | Hugging Face file bytes |
| UD-Q5_K_S | 77.8 GB | Hugging Face file bytes |
| UD-Q4_K_XL | 69.1 GB | Hugging Face file bytes |
| UD-Q4_K_M | 68.7 GB | Hugging Face file bytes |
| UD-Q4_K_S | 64.7 GB | Hugging Face file bytes |
| UD-IQ4_XS | 54.1 GB | Hugging Face file bytes |
| IQ4_NL | 55.1 GB | Hugging Face file bytes |
| UD-Q3_K_XL | 51.1 GB | Hugging Face file bytes |
| Q3_K_M | 50.6 GB | Hugging Face file bytes |
| UD-IQ3_S | 41.4 GB | Hugging Face file bytes |
| Q3_K_S | 46.2 GB | Hugging Face file bytes |
| UD-IQ3_XXS | 39.9 GB | Hugging Face file bytes |
| UD-Q2_K_XL | 37.4 GB | Hugging Face file bytes |
| UD-IQ2_XXS | 32.5 GB | Hugging Face file bytes |
| UD-IQ1_M | 30.1 GB | Hugging Face file bytes |
On common machines
| Device | Verdict | Quant | Speed |
|---|---|---|---|
| GeForce RTX 4060 Ti 8GB | Partial offload | UD-IQ1_M | 9.8–13 |
| GeForce RTX 3060 12GB | Partial offload | UD-IQ2_XXS | 10–14 |
| GeForce RTX 4060 Ti 16GB | Expert offload | UD-Q2_K_XL | 21–29 |
| GeForce RTX 4090 | Expert offload | UD-IQ3_XXS | 70–95 |
| GeForce RTX 5090 | Expert offload | UD-IQ3_XXS | 124–169 |
| Apple M4 Pro | Won't fit | UD-IQ1_M | — |
| Apple M5 | Won't fit | UD-IQ1_M | — |
| CPU only · DDR5-5600 | Won't fit | UD-IQ1_M | — |
Smallest devices that fit it fully
- NVIDIA A100 40GB — Tight fit, UD-IQ2_XXS, 131–179 tok/s
- NVIDIA L40S — Tight fit, UD-IQ3_S, 58–79 tok/s
- NVIDIA A40 — Tight fit, UD-IQ3_S, 46–63 tok/s
- RTX 6000 Ada — Tight fit, UD-IQ3_S, 64–87 tok/s
- RTX A6000 — Tight fit, UD-IQ3_S, 51–70 tok/s
- Apple M3 Max (40-core GPU) — Tight fit, UD-IQ2_XXS, 34–46 tok/s
Run it
Commands for the GeForce RTX 4060 Ti 16GB estimate (UD-Q2_K_XL).
llama.cpp
# Expert offload is approximate. llama.cpp can pin experts to CPU: # -ot "\\.ffn_.*_exps\\.=CPU" llama-server -hf unsloth/Mistral-Small-4-119B-2603-GGUF:UD-Q2_K_XL -c 8192 -ngl 99
Ollama
ollama run mistral
LM Studio
lms get unsloth/Mistral-Small-4-119B-2603-GGUF # Then load quant UD-Q2_K_XL