Published · Phase 3
Mistral Small 4 Hardware Requirements
Mistral Small 4's 119B total / 6.5B active MoE is a good 80–128 GB-class model. Use an 80–96 GB GPU for a reviewed NVFP4 build or DGX Spark for additional cache and precision headroom.
Editorial review: complete · Updated 2026-08-30
Direct answer
Direct answer guidance: Choose RTX PRO 6000 when the NVFP4 checkpoint plus cache fits below 85 GB and latency matters; choose Spark when 96 GB is tight or power/capacity matters more.
Model / quant variants
Model / quant variants guidance: Use the model-card-supported vLLM/SGLang path or NVIDIA-optimized serving image. llama.cpp depends on a mature compatible GGUF and MoE implementation.
Approximate weight footprint
Approximate weight footprint guidance: The official NVFP4 repository is roughly 70.8 GB. Treat 75–85 GB as a practical loaded budget and leave 10–20 GB more for KV cache, vision, and runtime workspaces.
Hardware tiers
Hardware tiers guidance: The official NVFP4 repository is roughly 70.8 GB. Treat 75–85 GB as a practical loaded budget and leave 10–20 GB more for KV cache, vision, and runtime workspaces. gpt-oss-120b is a text/reasoning alternative designed to fit 80 GB; Qwen3.6 is much smaller and easier; DeepSeek V4 is much larger.
Observed configurations
Observed configurations guidance: Only 6.5B parameters are active per token, so the model can be responsive once resident, but decode still depends on expert routing and memory bandwidth.
Context / concurrency effects
Context / concurrency effects guidance: The model advertises 256K context, but begin at 16K–32K and expand after measuring TTFT and cache use.
Performance expectations
Performance expectations guidance: Only 6.5B parameters are active per token, so the model can be responsive once resident, but decode still depends on expert routing and memory bandwidth.
Best-value configurations
Best-value configurations guidance: The model can justify professional memory if it outperforms 35B models on your workload. Otherwise a 32 GB consumer deployment is materially cheaper. Choose RTX PRO 6000 when the NVFP4 checkpoint plus cache fits below 85 GB and latency matters; choose Spark when 96 GB is tight or power/capacity matters more.
What will not fit / weak evidence
- Repository bytes do not equal peak runtime memory; packaging, scales, workspaces, graphs, and cache all add to the loaded process.
- Multimodal serving can add substantial encoder and cache allocations.
Planner CTA
Compare this recommendation against your model, context, concurrency, latency, and budget in the ComputeSage Planner.
Sources / methodology
Recommendations combine official specifications, public model/runtime documentation, adjacent benchmark observations, and clearly labeled engineering estimates. Review the ComputeSage evidence methodology.