DeepSeek V4 Flash (IQ2_XXS)
antirez's 2-bit community GGUF of DeepSeek V4 Flash — the 284B-parameter Mixture-of-Experts (~13B active per token) with 1M-token context and Compressed Sparse Attention. The routed experts are crushed to IQ2_XXS / Q2_K while attention projections, shared experts, router, embeddings, and output head stay at Q8_0 / F16, landing the whole model in ~81 GB — runnable on 128 GB-class machines versus ~156 GB for the native FP4 build. Note: these quants were produced for antirez's `ds4` inference engine; whether they load on stock llama depends on its DeepSeek-V4 support.
At a glance
- Parameters
- 284B
- Quantization
- IQ2_XXS
- Context window
- 1000K tokens
- Approx. size
- 86.7 GB
- Engines
- llama.cpp
- License
- MIT (open)
- Version
- v1.0.0 , released 2026-06-02
- 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)
- antirez/deepseek-v4-gguf ·
DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2.gguf· IQ2_XXS - Upstream
- https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash
- Manifest
- View on GitHub
Tags
- deepseek
- reasoning
- moe
- tools
- 2-bit
- iq2
- long-context
- large