DeepSeek-V4.1-Flash Quantized (FP8 weights / FP4 experts / FP4 KV cache)

DeepSeek-AI's DeepSeek-V4.1-Flash — a multimodal Mixture-of-Experts model with a 552B-parameter backbone that activates only 8B parameters per token at prefill and 16B at decode. It natively processes images and text and generates text autoregressively, at contexts up to 1,048,576 tokens (YARN, x16 from 65,536).

The release is built around KV-cache compression. A Causal Encoder-Decoder (CED) architecture — a 40-layer transformer split into a 20-layer causal encoder and a 20-layer decoder — projects the decoder's global KV cache from the final encoder states rather than per-layer, and Compressed Sparse Attention 2 (CSA2) assigns each attention layer one of three modes (Full / Reindex / Reuse) to share and reuse sparse-attention indices. Combined with FP4 main KV caching (E2M1, one E4M3 scale per 16 channels), the global KV cache drops to roughly 890 bytes per token — about 1/4 of DeepSeek-V4-Flash and ~437x smaller than DeepSeek-V1 — which makes it cheap on input-heavy agentic work. Additional components include Engram conditional memory (196B parameters, sparsely accessed) and DSpark speculative decoding. Each MoE layer has 1 shared + 384 routed experts, with 6 routed experts active per token.

The vision path is DeepSeek-ViT (trained from scratch, 2D-RoPE, 3x3 pixel-unshuffle downsampling) with a two-layer MLP projector, integrated from the start of language-model pre-training. The model was pre-trained on a 45T-token multimodal corpus, with context extended to 1M tokens late in training.

The instruct model exposes a continuously controllable reasoning effort (integer 1-100) trading inference cost for accuracy; published benchmarks use reasoning_effort=100. On agentic suites it is competitive with frontier closed models: Terminal-Bench 2.1 90.6 (top of DeepSeek's comparison table), DeepSWE v1.1 74.2, Codeforces 3471, CyberGym 88.1, AutomationBench 54.8, Agent's Last Exam 31.8.

Precision / quantization. This is a quantized deployment release, not full-precision BF16. Per config.json: FP8 weights (quant_method: fp8, 32x32 weight blocks, ue8m0 scale format) with FP4 experts (expert_dtype: fp4) and FP4 KV cache. The container dtype is bfloat16 but the weights themselves are quantized — which is why the full model is 510 GB rather than the ~1.1 TB a 552B backbone would occupy in BF16. There is no separate BF16 repo advertised; this appears to be the primary release form.

Chat template / inference. No Jinja chat template ships with this model. Prompt construction is done via the bundled encoding/ folder (a self-contained Python reference encoding.py with tests), or DeepSeek's deepseek-recipe Rust/Python toolkit for production. Weight-conversion and local-inference instructions are in the bundled inference/ folder; DeepSWE reproduction steps are in evaluation/. Recommended sampling: temperature=1.0, top_p=0.95 (or 1.0), 1M context, max_tokens >= 256K.

License note: MIT License (Copyright DeepSeek-AI), bundled as LICENSE. Fully permissive — use, copy, modify, merge, publish, distribute, sublicense, and sell, for both the repository and the model weights, with only the standard notice-retention condition. No revenue gate, no acceptable-use restrictions, no Model-as-a-Service clause — notably more permissive than the GLM-5.x releases.

What's in the torrent. Unlike a weights-only mirror, this repo bundles runnable code and docs (89 files, 32 MiB pieces): 49 model-*.safetensors shards of FP8/FP4 quantized weights plus model.safetensors.index.json; config.json, tokenizer.json, and tokenizer_config.json (129,280-token vocab, no chat template); the encoding/, inference/, and evaluation/ folders; assets/ benchmark figures; the DeepSeek-V4.1 technical report PDF (1.8 MB); README.md, LICENSE, and .gitattributes. All 89 files were verified against Hugging Face's published hashes before the torrent was built; the full local manifest is in DeepSeek-V4.1-Flash.sha256. Verify your download against those checksums.

Caveats worth knowing. This is a quantized release (FP8 weights + FP4 experts/KV), not BF16 — that is the model's native form here, half a terabyte rather than a truncation, and no BF16 repo is published. There is no chat template, so integrators must use the bundled encoding/ reference or deepseek-recipe; a naive apply_chat_template will not work. It is not a single-GPU model — a 552B backbone across 49 shards is a multi-node or heavily-offloaded deployment even quantized, and the whole 510 GB is required (there is no useful subset). It is multimodal: a text-only serving setup ignores half of what the model was built for. And the architecture is novel — CED, CSA2, Engram, and DSpark are non-standard, so stock transformers/vLLM/SGLang support may lag; check the bundled inference/ folder and deepseek-recipe for the currently supported path before promising it runs out of the box.

Size and file count are taken from the torrent metadata itself. Created and permanently seeded by drforbin.ai — this is the site's own mirror of the official Hugging Face DeepSeek-V4.1-Flash release.