license: other license_name: deepseek-license base_model:

• deepseek-ai/DeepSeek-V4.1-Flash
• dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8 tags:
• deepseek-v4.1
• moe
• affinity
• rocm
• rdna4
• quantized
• uncensored library_name: affinity pipeline_tag: text-generation

DeepSeek-V4.1-Flash-UNCENSORED — affinity container (2.875 bpw)

A quantised affinity container of the uncensored DeepSeek-V4.1-Flash, made to run on two AMD Radeon AI PRO R9700 (gfx1201) GPUs. It loads only in the affinity engine (the C++/HIP inference engine by Yoshi Exeler) with the V4.1 port — not in llama.cpp, vLLM or SGLang.

Full setup, engine patches, deployment and measurements live in the companion GitHub repo: https://github.com/drwolfen/deepseek-v4.1-flash-r9700-affinity

Public model files hosted for free on the Hub as 5 parts (the main container exceeds the per-file limit). Nothing here is an official DeepSeek or affinity release; it is a personal quantisation of an abliterated community checkpoint.

Files

Reassemble:

cat model-uc.aff.part-* > model-uc.aff

Then verify the pieces you downloaded match SHA256SUMS.parts (which was produced from these parts): sha256sum model-uc.aff.part-* and compare.

What this is

• Architecture: DeepSeek-V4.1-Flash — 40 layers, 384 routed experts (top-6), hidden 5120, 129280 vocab, two-level sparse-attention indexer, engram n-gram layers, vision tower (DeepSeek ViT).
• Quantisation: only the routed experts are quantised — 2.875 bpw with affinity's codebook quant (2 variants × 1024 entries), imatrix-weighted. Dense weights stay in the checkpoint's FP8 (6.2 GiB), embedding/head in bf16. Routed experts ≈ 93 GiB of the 191.32 GiB total.
• Uncensored: the source is dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8 (DPO/refusal-ablated). An alternative build from pqhaz/DeepSeek-V4.1-Flash-heretic (directional ablation, lower KL 0.210 vs 0.333) can be quantised the same way; this container is the -uc one.

How it runs (speed, hardware)

Measured on the machine below, single stream, 512k-context config, DSpark speculative decoding on, thinking on:

• Decode 18–19.6 tok/s long-context reasoning (median ≈ 18.9), 24–25 tok/s on high-acceptance bursts; draft acceptance 54–93 %.
• Prefill ~1000 tok/s at 16k–48k (prefix-cache hit 98.7 %).
• Hardware: 2 × Radeon AI PRO R9700 (gfx1201, 32 GiB), Xeon Platinum 8160, 754 GiB RAM. Minimum ~160 GiB RAM for this configuration: 3274/15360 experts are resident in VRAM, the other 12086 are read from a 143.5 GiB host pool over PCIe. Decode is PCIe-bandwidth bound; the cards link at Gen5 but the host's switch uplink is Gen3 ×16 (PCIe-3.0 CPU), which saturates — a Gen4 ×16 host is the largest available speedup (~1.6×).

Engine changes and optimisations

This container is served by a fork of affinity that adds (full detail + commit list in the GitHub repo):

• DeepSeek ViT on the CPU, wired into prefill and the chat endpoint, with a weight-stationary GEMM + AVX-512 attention (546×546 image: 13.2 s → 3.25 s; 1024×1024: minutes → 38 s) and corrected 2D-RoPE / aligner / preprocessing (embedding cosine vs reference 0.31 → 0.85).
• Compressed KV sized per kv source (KV 1016 → 100 MiB/card) and indexer mask tiling (prefill arena 2068 → 1426 MiB/card) — both raise how many experts stay resident.
• Marker-less ("bare") DSML tool-call recovery and stop-at-block-close, which stops the agentic runaway/confabulation on a long context (a 117k-token replay: 4360 → 195 completion tokens).
• Learned placement heat-plane persistence across restarts (no re-learn from zero).
• DSpark confidence head and opt-in backoff/adaptive-width knobs for speculative decoding.
• On-GPU imatrix calibration (the engine can emit its own imatrix from the model's own activations).
• AFF_NOBUDGET_MAX_TOKENS: bounds the default output when a client sends no max_tokens, so a gateway reasoning loop cannot hold the single slot for tens of minutes.

Usage

docker run --device=/dev/kfd --device=/dev/dri --group-add video --ipc=host \
  -v $PWD:/models affinity:v41-vision \
    -m /models/model-uc.aff --tokenizer /models/tokenizer.json \
    --dspark on --dspark-model /models/model-uc.dspark.aff \
    --engram /models/model-uc.engram --engram-tables /models \
    --kv-size 524288 --kv-commit 262144 \
    --host-pool-mib 200000 --placement-strategy demand \
    --prefill-chunk-size 2816 --default-effort 75 --max-tokens-floor 0

Credits and license

Engine: StillDeadcode/affinity (MIT), by Yoshi Exeler. Optimization ports: Luce-Org/lucebox · lucebox.com (Apache-2.0, Lucebox AI) — the bounded-replay, adaptive-verify-width and protected-outlier-expert optimizations in this engine fork are ported from Lucebox's ds41-gorgon work, and the port's quality gate reuses their fineweb-edu eval corpus; thanks to Lucebox. Base model: deepseek-ai/DeepSeek-V4.1-Flash (DeepSeek license). Uncensored weights: dealignai. GGUF/imatrix: Adolphsson. This quantisation and the engine fork are derivative work; the DeepSeek license governs the weights. The transformation of the affinity engine into this R9700-tweaked build was done with the help of DeepSeek-V4.1-Flash and Gemini 4 Argon. Released for use in environments you own or are authorised to test.