DeepSeek V4 Flash (FP4)
DeepSeek V4 Flash in its native FP4+FP8 mixed-precision build — 284B-parameter Mixture-of-Experts (~13B active per token) with 1M-token context, Compressed Sparse Attention, and configurable Non-think / Think-High / Think-Max reasoning modes. The FP4 weights run in ~156 GB on llama via the `nsparks` GGUF (the FP4-FP8 native mix the upstream was trained at), and as a 4-bit MLX directory on Apple Silicon. Frontier-tier on-device reasoning for workstations and Mac Studio-class hardware — needs ~200 GB of usable memory.
At a glance
- Parameters
- 284B
- Quantization
- FP4-FP8
- Context window
- 1000K tokens
- Approx. size
- 156.1 GB
- Engines
- llama.cpp, MLX
- License
- MIT (open)
- Version
- v1.0.0 , released 2026-05-17
- Family
- deepseek
Tuning
Defaults a gezel client applies out of the box, from real evaluation runs against this exact quantization.
- sampling.maxTokens
16384- reasoning.thinkingBudget
4096- reasoning.enableThinking
true
Model behaviors
Client-side behavior modules the catalog enables for this model (reasoning-tag handling, fabrication detection, prompt shaping).
- reasoning.strip-think-tags
- prompt.private-reasoning-guidance
- prompt.tool-cookbook-condensed
- mcp.compact-tool-schemas
- fabrication.detect-past-tense-no-tools
Sources
- llama.cpp (GGUF)
- nsparks/DeepSeek-V4-Flash-FP4-FP8-GGUF ·
DeepSeek-V4-Flash-FP4-FP8-native.gguf· FP4-FP8 - MLX
- mlx-community/deepseek-ai-DeepSeek-V4-Flash-4bit
- Upstream
- https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash
- Manifest
- View on GitHub
Tags
- deepseek
- reasoning
- moe
- tools
- fp4
- long-context
- large