Mistral Small 4 119B-A6.5B

Mistral's Apache-2.0 multimodal MoE unifies instruct, reasoning, and coding modes. It has 119B total parameters, 6.5B active, 128 experts with 4 routed per token, MLA attention, and an official 256K context window.

Mistral AIMixture of expertsgeneralreasoning17 machines can run it
Run locally →
Total parameters
119.401B
Official spec sets memory need
Active parameters
6.5B
Official spec sets decode speed
Context
256K
Official spec tokens
4-bit weights
65 GB
Calculated before KV cache
This is a sparse mixture of experts. All 119.401B of weights must be available to the runtime, but only 6.5B are read per generated token. Ordinary runtimes keep the full quantized model in memory; a runtime-specific SSD-streaming artifact can retain a smaller working set and fetch expert data on demand, trading speed for capacity.
Model specification
Publisher
Mistral AI Official spec
Architecture
Sparse mixture of experts Official spec
Total parameters
119.401B Official spec
Active parameters
6.5B per token Official spec
Context length
262,144 tokens Official spec
Attention
MLA · 32 KV heads × 128 dim × 36 layers Official spec
Licence
Apache 2.0 Official spec source
Specification confidence
High · verified 2026-09-21 Official spec source
Released
16 Mar 2026 Official spec
Official source
Quantizations and memory
QuantizationFormatBits/weightResidentDownloadQuality kept
MLX 4-bitmlx4.565.1 GB65.1 GB98.0%
Q4_K_Mdefaultgguf4.8569.4 GB69.4 GB98.5%
Q5_K_Mgguf5.6981.5 GB81.5 GB99.3%
Q8_0gguf8.5120.5 GB120.5 GB99.9%
BF16safetensors16222.4 GB222.4 GB100.0%
Generic weight sizes are computed from the parameter count and bits per weight plus a format-specific overhead. Runtime-specific artifacts use their published resident and download footprints; streamed models can therefore require much more disk than memory. Quality retention is an assumption, not a measured evaluation.
Coding & quality benchmarksCompare coding results →

No published benchmark results have been imported for this model yet. This is missing evidence, not a score of zero.

We show source metrics rather than deriving one opaque quality number. Different benchmarks measure genuinely different things, and collapsing them into a single score would hide exactly the disagreements worth seeing.
Recommendations
Cheapest that can run it
~4.5–6.5 t/s · $1,531
Lowest purchase price among configurations where the model fits at some quantization in our catalogue. Speed is not considered.
Cheapest above 20 t/s
~18.3–26.4 t/s · $1,711
20 tokens/second is roughly the point at which generation keeps pace with reading. Below it, interactive use feels like waiting.
Cheapest above 40 t/s
~73.8–106 t/s · $4,323
40 tokens/second is the threshold most people describe as comfortable for coding agents, where output arrives faster than you can review it.
Fastest with real measurements
Nothing in the database qualifies.
Highest throughput among configurations with an actual published measurement rather than our estimate.
Best throughput per purchase-price unit
~152–218 t/s · $5,945
Highest decode tokens/second per 1,000 units of the displayed purchase currency. Ignores running costs and resale — see the economics section for the full picture.
Hardware that runs Mistral Small 4 119B-A6.5B0 measured, 17 estimated
17 of 17 rows
HardwareQuantMemoryDecodePrefillContextPriceFitConfidence
NVIDIA DGX H200 (8x H200, 1,128 GB)
NVIDIA · 1,128 GB · 38,400 GB/s
Q8_0
recommended
128.1 GB~483–695 t/s~96040–199470 t/s256K$359,430ComfortableEstimated
Quad RTX 5090 workstation (4x 32 GB)
NVIDIA · 128 GB · 7,168 GB/s
Q5_K_M
recommended
86.3 GB~261–375 t/s~3600–7470 t/s256K$12,611ComfortableEstimated
RTX PRO 6000 Blackwell workstation (96 GB)
NVIDIA · 96 GB · 1,792 GB/s
Q4_K_M
recommended
72.7 GB~192–276 t/s~1170–2430 t/s256K$10,359FitsEstimated
RTX PRO 6000 Max-Q workstation (96 GB, 300 W)
NVIDIA · 96 GB · 1,792 GB/s
Q4_K_M
recommended
72.7 GB~192–276 t/s~957–1990 t/s256K$10,809FitsEstimated
Lenovo ThinkStation PX (4x RTX PRO 6000, 384 GB)
Lenovo · 384 GB · 7,168 GB/s
Q8_0
recommended
126.5 GB~185–266 t/s~4610–9580 t/s256K$58,553ComfortableEstimated
Mac Studio M5 Ultra 96 GB
Apple · 96 GB · 1,200 GB/s
Q4_K_M
recommended
72.7 GB~152–218 t/s~833–1730 t/s256K$5,945FitsEstimated
Dell Precision 7960 Rack (2x RTX PRO 6000, 192 GB)
Dell · 192 GB · 3,584 GB/s
Q8_0
recommended
125.7 GB~139–200 t/s~2310–4790 t/s256K$36,663ComfortableEstimated
NVIDIA DGX Station GB300 (748 GB)
NVIDIA · 748 GB · 7,100 GB/s
Q8_0
recommended
125.3 GB~95–137 t/s~12830–26650 t/s256K$90,082ComfortableEstimated
Mac Studio M5 Ultra 512 GB
Apple · 512 GB · 1,200 GB/s
Q8_0
recommended
125.3 GB~94–135 t/s~1040–2160 t/s256K$11,350ComfortableEstimated
Mac Studio M5 Ultra 256 GB
Apple · 256 GB · 1,200 GB/s
Q8_0
recommended
125.3 GB~94–135 t/s~1040–2160 t/s256K$8,647ComfortableEstimated
Mac Studio M5 Max 128 GB
Apple · 128 GB · 614 GB/s
Q5_K_M
recommended
85.1 GB~73.8–106 t/s~521–1080 t/s256K$4,323ComfortableEstimated
Mac Studio M3 Ultra 512 GB
Apple · 512 GB · 819 GB/s
Q8_0
recommended
125.3 GB~54.6–78.6 t/s~251–522 t/s256K$8,287ComfortableEstimated
Mac Studio M3 Ultra 256 GB
Apple · 256 GB · 819 GB/s
Q8_0
recommended
125.3 GB~54.6–78.6 t/s~251–522 t/s256K$5,674ComfortableEstimated
NVIDIA DGX Spark 128 GB
NVIDIA · 128 GB · 273 GB/s
Q4_K_M
recommended
72.7 GB~30.5–44 t/s~448–931 t/s256K$3,873ComfortableEstimated
Framework Desktop (Ryzen AI Max+ 395, 128 GB)
AMD · 128 GB · 256 GB/s
Q4_K_M
recommended
72.7 GB~21–30.2 t/s~124–257 t/s256K$2,161ComfortableEstimated
GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB)
AMD · 128 GB · 256 GB/s
Q4_K_M
recommended
72.7 GB~18.3–26.4 t/s~124–257 t/s256K$1,711ComfortableEstimated
CPU-only workstation (Ryzen 9950X, 192 GB DDR5)
Generic · 192 GB · 90 GB/s
Q8_0
recommended
125.3 GB~4.5–6.5 t/s~14.9–31 t/s256K$1,531ComfortableEstimated
Submit a benchmarkContributions are reviewed before publication.
Similar models