Skip to content
Other

DeepSeek-V4.1-Flash-1017-radiance

Running DeepSeek-V4.1 on AMD GPUs

2 min readupdated

Developer aselzer converted the DeepSeek-V4.1-Flash-1017 large language model into the .rad format for the AMD-focused radiance inference engine. This build is designed to reproduce high-load batching benchmarks from the BatchSeek project (handling 512 concurrent requests). Inside the main file, the original experts are preserved in MXFP4 encoding alongside the remaining weights in BF16/F32, while Engram tables are added in IQ4_NL compression. It is built for enthusiasts and researchers running large MoE models on clusters of consumer AMD cards without NVIDIA hardware.

Key features

  • radiance format (.rad): The main 416 GB file was compiled using the radiance converter (commit bd65d4e) with DeepSeek-V4.1 architecture support.
  • Chunked downloads: The model is split into 8 parts of roughly 52 GB each, complete with checksum verification via SHA256SUMS.
  • Two Engram table modes: Built-in quantized IQ4_NL tables are used by default, though you can attach the original FP8 shards (+203 GB) with no performance penalty.
  • Permissive MIT license: The weights belong to DeepSeek; this repository only changes the storage format.

What you can do

  • Reproduce BatchSeek benchmarks: Test parallel generation across 512 streams using the serve-batch.sh script.
  • Deploy a local AMD GPU server: Run the model across a multi-GPU Radeon setup (tested on 4× RX 9070 XT) paired with high-capacity system RAM.
  • Compare quantization accuracy: Benchmark the built-in IQ4_NL tables against the original FP8 shards on your own workloads.

Example

./serve-batch.sh --model model.rad --silu-table silu_table_gfx1201.bin --streams 512 \
  --engram-file engram-fp8/model-00047-of-00048.safetensors,engram-fp8/model-00048-of-00048.safetensors
Read moreDetails
BIRD 256 benchmark: 170 s (NLL 0.688) with the original FP8 tables vs. 172 s (NLL 0.692) with the built-in IQ4_NL tables.
Test rig: 4× AMD RX 9070 XT (16 GB VRAM each, gfx1201 architecture), EPYC 7443 CPU, and 256 GB DDR4.
File sizes: 416 GB for model.rad, plus 203 GB if using the optional engram-fp8 shards.
Discuss Telegram 0

Will it run?

Won't run

Not enough memory even for 4-bit. Below are GPUs you can rent.

  • BF16weights 189 GB · ready files209 GB
  • FP8 / INT8recommendedweights 95 GB105 GB
  • 4-bitweights 53 GB59 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: Won't fit any home GPU — renting is cheaper.

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # aselzer/DeepSeek-V4.1-Flash-1017-radiance
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download aselzer/DeepSeek-V4.1-Flash-1017-radiance --local-dir ~/models/DeepSeek-V4.1-Flash-1017-radiance
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of DeepSeek-V4.1-Flash-1017 (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 9All on Hugging Face
LoRAs and adapters 2All on Hugging Face
Merges 1All on Hugging Face

About the model

aselzer/DeepSeek-V4.1-Flash-1017-radiance
Author
aselzer
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments