DeepSeek-V4.1-Flash-1017 open weights: 763B parameters, a 510 GB download, and what it takes to run locally

The model DeepSeek moved behind deepseek-flash on October 17 is also downloadable. deepseek-ai/DeepSeek-V4.1-Flash-1017 went up on Hugging Face on October 17, 2026. The model card says “this repository and the model weights are licensed under the MIT License.” If you only call the API, our 1017 API update note covers the benchmark table and pricing. This post answers the local question: how big is it, and what can actually load it today?

The checkpoint in numbers

From the Hugging Face repo metadata and config.json:

• 763,205,315,794 parameters, stored in 48 safetensors shards totalling about 510 GB.
• Mixed precision as shipped: FP8 block-quantized weights, with routed experts in FP4 (expert_dtype: fp4 in the quantization config).
• MoE layout: 40 layers, 384 routed experts plus 1 shared expert, 6 experts per token.
• Context: max_position_embeddings is 1,048,576. The card recommends temperature 1.0, top_p 1.0, a 1M-token context window and max_tokens of at least 256K.
• Multimodal: the pipeline tag is image-text-to-text, and the repo includes a vision encoder.

The mlx-community oQ4e card breaks the total down further: a 551.9B language backbone, 196.6B in Engram lookup tables, a 0.49B vision encoder and a 14.2B MTP head. DeepSeek’s own card does not publish an active-parameter count.

No chat template: use the encoder or deepseek-recipe

This release does not include a Jinja-format chat template. Tools that build prompts from tokenizer_config.json won’t format conversations correctly out of the box. DeepSeek ships two alternatives instead:

1. encoding/encoding.py, a self-contained Python reference with test cases for multi-turn chat, tool calls, thinking mode, numeric reasoning effort and interleaved images.
2. deepseek-recipe, MIT-licensed Rust libraries with Python bindings. They convert Messages, Chat Completions and Responses API requests into DeepSeek V4/V4.1 prompts or token IDs, and parse output back, including streamed tool calls.

DeepSeek’s reference runner assumes 8 GPUs

The bundled inference/ folder is described as “a readable reference implementation rather than a production serving engine.” It needs torch>=2.10.0 and tilelang==0.1.12. You first convert the weights into one file per tensor-parallel rank, then launch with torchrun:

python convert.py --hf-ckpt-path "${HF_CKPT_PATH}" --save-path "${SAVE_PATH}" \
  --model-parallel 8 --expert-dtype fp4 --tokenizer-path "${HF_CKPT_PATH}"

torchrun --nproc-per-node 8 generate.py --ckpt-path "${CKPT_PATH}" \
  --config config.json --interactive --temperature 0.6

Multi-node runs take the usual torchrun --nnodes / --node-rank arguments.

Community quants: GGUF and MLX

Several unofficial quants appeared within hours. Sizes below are taken from each repo:

Repo	Format	Size
mrfakename GGUF	Q2_K / Q3_K_M / Q4_K_M / Q8_0	246 / 323 / 414 / 473 GiB
mlx-community oQ4e-mtp	MLX mixed precision	402.4 GiB
mlx-community oQ3e-mtp / oQ2.5e-mtp	MLX mixed precision	about 380 / 343 GB

Two caveats come from the GGUF card. First, mainline llama.cpp does not support DeepSeek V4.1 yet: the files need the branch from llama.cpp PR #28696, which was still open on October 17. Second, they are text only, because that branch can’t convert the vision tower yet. The card suggests keeping the Engram tables on CPU with -ot "engram_embd.weight=CPU". Q4_K_M saves less than you might expect, because the routed experts are already about 4-bit.

Even the smallest quant needs roughly a quarter-terabyte of memory. That puts it in the same multi-machine class as Meta’s 816B model, so the hardware options in our Muse Spark Open local guide and the Mac Studio RDMA cluster write-up apply here too. For most builders, the deepseek-flash API in our V4 Flash tutorial remains the practical path.

Sources: DeepSeek-V4.1-Flash-1017 model card, inference README, deepseek-recipe, mrfakename GGUF card, mlx-community oQ4e card, llama.cpp PR #28696.