DeepSeek V4 Flash — GGUF for ds4

This quants are specific for the DS4 inference engine. They may work with other inference engines or not (they should, but not the MTP model which requires a specific loader).

https://github.com/antirez/ds4

Files

File	Size	Routed experts (ffn_{gate,up,down}_exps)	Everything else
DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2.gguf	80.8 GiB	IQ2_XXS (gate, up) + Q2_K (down)	Q8_0 attn proj / shared experts / output, F16 router + embed + indexer + compressor + HC, F32 norms / sinks / bias
DeepSeek-V4-Flash-Q4KExperts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2.gguf	153.3 GiB	Q4_K (all three)	same as above
DeepSeek-V4-Flash-MTP-Q4K-Q8_0-F32.gguf	3.6 GiB	MTP / speculative-decoding support (optional, not standalone).

Use q2 on 128 GB Mac machines, q4 on machines with ≥ 256 GB RAM, pair either with MTP for optional speculative decoding.

Quantization recipe

The filename is the spec. In detail, for the q2 file:

Tensor class	Quant	Notes
blk.*.ffn_gate_exps, blk.*.ffn_up_exps	IQ2_XXS	routed-expert up/gate
blk.*.ffn_down_exps	Q2_K	routed-expert down (K-quant for quality)
blk.*.ffn_{gate,up,down}_shexp	Q8_0	shared experts
blk.*.attn_q_a, attn_q_b, attn_kv, attn_output_a, attn_output_b	Q8_0	all attention projections (MLA + low-rank output)
output.weight	Q8_0	output head
token_embd.weight	F16	input embedding
blk.*.ffn_gate_inp (router)	F16	learned router
blk.*.exp_probs_b (router bias), blk.*.attn_sinks, all *_norm.weight	F32
blk.*.ffn_gate_tid2eid	I32	hash-routing tables (first 3 layers only)
blk.*.attn_compressor_*, blk.*.indexer_*, blk.*.hc_*, blk.*.output_hc_*	F16 / F32	DSv4-specific auxiliary blocks

For the q4 file, only the three routed-expert classes change to Q4_K. Everything else is byte-for-byte identical to the q2 recipe.

The motivation behind the asymmetry: the routed experts are the majority of the parameter count but each individual expert handles only a fraction of tokens, so aggressive quantization on them costs less in average quality than the same treatment of router, projections, or shared experts. Keeping the decision-making components at Q8_0 preserves model behavior; crushing the experts buys the size.

Usage

git clone https://github.com/antirez/ds4
cd ds4
./download_model.sh q2     # 128 GB RAM machines
./download_model.sh q4     # >= 256 GB RAM machines
./download_model.sh mtp    # optional MTP / speculative decoding
make

./ds4 -p "Explain Redis streams in one paragraph."
./ds4-server --ctx 100000 --kv-disk-dir /tmp/ds4-kv --kv-disk-space-mb 8192

The download_model.sh script fetches from this repository, resumes partial downloads, and points ./ds4flash.gguf at the selected variant.

License

MIT. The base model copyright is held by DeepSeek; the GGUFs are redistributed under the base model's release terms.