← All modelsMODEL CHECK

Can I run Ministral 3 8B?

Ministral 3 8B by Mistral AI needs around 8 GB of RAM at the recommended 4-bit quantization (4.9 GB download). Your hardware is checked below β€” instantly, nothing leaves your browser. Expect roughly ~63 tok/s on a NVIDIA RTX 3060 12GB.

Reading your hardware signals…

Real-world notes

Ministral 3 8B is Mistral's late-2025 take on the small, do-everything local model, and it handles both chat and vision in the same 8B package. At a 4-bit quant it comes in around 4.9 GB, so it slots onto an 8 GB GPU with room to spare and runs happily within unified memory on any Apple Silicon Mac. The 8 GB minimum RAM figure is honest for the weights alone, which makes this a realistic pick for a laptop or a cheap second-hand GPU rather than something you need a workstation to host.

In daily use it feels quick. On an RTX 3060 you can expect around 63 tokens per second at 4-bit, a 4090 pushes that to roughly 177, and an M-series Max sits near 72, all faster than you can read a streaming reply. The headline is the 256K context window, but treat it as a ceiling, not a default. At 128K of context the full memory footprint climbs to about 21.7 GB, well past what an 8 GB card holds, so keep working context to a few thousand tokens unless you have the VRAM to back the long window.

Against its siblings it splits the difference: Mistral Nemo 12B generally has more headroom for harder reasoning if you can afford the larger model, while Mistral 7B is the leaner fallback when memory is tight. Ministral 3 8B's standout trait is that vision sits in the same small footprint, so you get image understanding without jumping to a heavier multimodal model. It ships under Apache 2.0, which means you can use it commercially and in production with no provider-specific strings attached, a genuinely clean license for a model this capable.

Specifications

Parameters8B
Context window256K tokens
ProviderMistral AI
LicenseApache 2.0
Released2025-12
Best forChat, Vision

Size by quantization

QuantizationBits/weightDownloadMin RAMQuality
Q2_K3.353.4 GB6 GBNoticeable loss
Q4_K_MRecommended4.854.9 GB8 GBRecommended
Q5_K_M5.655.7 GB12 GBHigh
Q8_08.58.5 GB16 GBNear-original
F161616.0 GB24 GBOriginal

Sizes are estimates from parameter count Γ— bits per weight; real GGUF builds vary slightly. Β· Data updated: 2026-06-11 Β· How we calculate these numbers β†’

Memory needed by context length

ContextKV cache (est.)Total memory (Q4)
4K tokens~0.5 GB~5.4 GB
8K tokens~1.0 GB~5.9 GB
32K tokens~4.2 GB~9.1 GB
128K tokens~16.8 GB~21.7 GB

The KV cache grows with context length β€” a model that fits at 4K can run out of memory at 32K. Estimates assume an FP16 cache with grouped-query attention; actual usage varies by runtime.

Estimated speed by hardware

HardwareBandwidth~Speed
NVIDIA RTX 3060 12GB360 GB/s~63 tok/s
NVIDIA RTX 4090 24GB1008 GB/s~177 tok/s
Apple M-series (base)100 GB/s~18 tok/s
Apple M-series Pro270 GB/s~47 tok/s
Apple M-series Max410 GB/s~72 tok/s
CPU only (dual-channel DDR5)60 GB/s~11 tok/s

Token generation is memory-bandwidth bound: tok/s β‰ˆ bandwidth Γ— 0.85 Γ· model size at Q4. Real-world numbers vary by runtime and context length.

Run it locally

The easiest path is Ollama β€” one command and you're chatting:

ollama run ministral-3:8b

Frequently asked questions

Ministral 3 8B System Requirements β€” Can I Run It Locally?