README.md
| 1 | --- |
| 2 | license: mit |
| 3 | library_name: gguf |
| 4 | pipeline_tag: text-generation |
| 5 | base_model: deepseek-ai/DeepSeek-V4-Flash |
| 6 | base_model_relation: quantized |
| 7 | quantized_by: antirez |
| 8 | language: |
| 9 | - en |
| 10 | tags: |
| 11 | - gguf |
| 12 | - quantized |
| 13 | - deepseek |
| 14 | - deepseek-v4 |
| 15 | - deepseek-v4-flash |
| 16 | - moe |
| 17 | - mixture-of-experts |
| 18 | - 2-bit |
| 19 | - 4-bit |
| 20 | - iq2_xxs |
| 21 | - q2_k |
| 22 | - q4_k |
| 23 | - ds4 |
| 24 | - apple-silicon |
| 25 | - metal |
| 26 | --- |
| 27 | |
| 28 | # DeepSeek V4 Flash — GGUF for ds4 |
| 29 | |
| 30 | 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). |
| 31 | |
| 32 | https://github.com/antirez/ds4 |
| 33 | |
| 34 | ## Files |
| 35 | |
| 36 | | File | Size | Routed experts (`ffn_{gate,up,down}_exps`) | Everything else | |
| 37 | |---|---:|---|---| |
| 38 | | `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 | |
| 39 | | `DeepSeek-V4-Flash-Q4KExperts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2.gguf` | 153.3 GiB | `Q4_K` (all three) | same as above | |
| 40 | | `DeepSeek-V4-Flash-MTP-Q4K-Q8_0-F32.gguf` | 3.6 GiB | MTP / speculative-decoding support (optional, not standalone). | | |
| 41 | |
| 42 | Use **q2** on 128 GB Mac machines, **q4** on machines with ≥ 256 GB RAM, pair either with **MTP** for optional speculative decoding. |
| 43 | |
| 44 | ## Quantization recipe |
| 45 | |
| 46 | The filename is the spec. In detail, for the **q2** file: |
| 47 | |
| 48 | | Tensor class | Quant | Notes | |
| 49 | |---|---|---| |
| 50 | | `blk.*.ffn_gate_exps`, `blk.*.ffn_up_exps` | **`IQ2_XXS`** | routed-expert up/gate | |
| 51 | | `blk.*.ffn_down_exps` | **`Q2_K`** | routed-expert down (K-quant for quality) | |
| 52 | | `blk.*.ffn_{gate,up,down}_shexp` | `Q8_0` | shared experts | |
| 53 | | `blk.*.attn_q_a`, `attn_q_b`, `attn_kv`, `attn_output_a`, `attn_output_b` | `Q8_0` | all attention projections (MLA + low-rank output) | |
| 54 | | `output.weight` | `Q8_0` | output head | |
| 55 | | `token_embd.weight` | `F16` | input embedding | |
| 56 | | `blk.*.ffn_gate_inp` (router) | `F16` | learned router | |
| 57 | | `blk.*.exp_probs_b` (router bias), `blk.*.attn_sinks`, all `*_norm.weight` | `F32` | | |
| 58 | | `blk.*.ffn_gate_tid2eid` | `I32` | hash-routing tables (first 3 layers only) | |
| 59 | | `blk.*.attn_compressor_*`, `blk.*.indexer_*`, `blk.*.hc_*`, `blk.*.output_hc_*` | `F16` / `F32` | DSv4-specific auxiliary blocks | |
| 60 | |
| 61 | 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. |
| 62 | |
| 63 | 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. |
| 64 | |
| 65 | ## Usage |
| 66 | |
| 67 | ```bash |
| 68 | git clone https://github.com/antirez/ds4 |
| 69 | cd ds4 |
| 70 | ./download_model.sh q2 # 128 GB RAM machines |
| 71 | ./download_model.sh q4 # >= 256 GB RAM machines |
| 72 | ./download_model.sh mtp # optional MTP / speculative decoding |
| 73 | make |
| 74 | |
| 75 | ./ds4 -p "Explain Redis streams in one paragraph." |
| 76 | ./ds4-server --ctx 100000 --kv-disk-dir /tmp/ds4-kv --kv-disk-space-mb 8192 |
| 77 | ``` |
| 78 | |
| 79 | The `download_model.sh` script fetches from this repository, resumes partial downloads, and points `./ds4flash.gguf` at the selected variant. |
| 80 | |
| 81 | ## License |
| 82 | |
| 83 | MIT. The base model copyright is held by DeepSeek; the GGUFs are redistributed under the base model's release terms. |
| 84 | |