license: mit
library_name: radiance
pipeline_tag: image-text-to-text
tags:
- radiance
- rocm
- amd
- rdna4
- deepseek
- deepseek-v4.1
- uncensored
- abliterated
- tp4
- iq2_xxs
- vision
base_model: - deepseek-ai/DeepSeek-V4.1-Flash-1017
DeepSeek-V4.1-Flash-1017 — IQ2_XXS vision, uncensored
DeepSeek-V4.1-Flash-1017 as a single .rad container for the radiance inference engine (AMD RDNA4, ROCm), with the release's 32-layer vision encoder kept as bf16 so images are accepted through the chat API. The refusal direction is edited out of 18 attention-output projections.
Weights are in this repo.
Container size: 259.09 GiB (278,197,639,056 bytes).
| Container | 259.09 GiB / 278,197,639,056 bytes |
| Files | deepseek-v4.1-flash-1017-iq2xxs-vision-uncensored.rad.part-00 … .part-05 — six parts (45 GiB each, last 34.09 GiB; the Hub's per-file limit is 50 GB) |
| Reassemble | cat deepseek-v4.1-flash-1017-iq2xxs-vision-uncensored.rad.part-0[0-5] > deepseek-v4.1-flash-1017-iq2xxs-vision-uncensored.rad (sha256 77c5d2e021af59e374e368a8b1f05d145bb4f1a2fe6f687c64e39236b167cfb0) |
| Base | DeepSeek-V4.1-Flash-1017 (fp8 weights, fp4 routed experts; MIT) |
| Routed experts | IQ2_XXS (iq2_xxs, ≈2.06 bits/weight) on every routed-expert layer |
| Attention | q6_K on layers 0–26, q5_K on layers 27–39; DSpark layers 40–42 bf16 |
| Shared experts | q6_K and q8_0 |
| Engram table | iq4_nl, 103 GiB, rebuilt from the release tokenizer; the release ships no Engram table |
| Embeddings / output | q6_K |
| Vision encoder | bf16; 32 blocks, hidden 1024, 16 heads of width 64, patch 14, 2D RoPE, 3×3 aligner; 0.904 GiB as stored |
| DSpark drafter | MXFP4 experts as released, 6.72 GiB |
| Abliteration | attn_output_b of layers 11–28, alpha 0.75; the other 18 layers stay stock |
| Norms / routers | f32 |
| Measured | 56.9 tok/s decode (256 tokens), 24,013-token prompt TTFT 7.03 s on four 32 GB RDNA4 cards, TP4, n-gram 5 speculation |
Run it today: upstream
maindoes not carry thearch/deepseek41plugin yet. Clone thedeepseek41-wipbranch ofbakon/radianceand build it. The download, join, build and serve commands are in RUNNING.md.
Quantisation
This is a per-tensor mix, not one global type. The routed experts were re-quantised from the official checkpoint with the IQ2_XXS search; the attention, shared experts, embedding and output planes stay at their higher-width quantisers; the vision encoder is bf16. The 259 vision tensors, four aligner tensors, three delimiter vectors and 43 bias_vl vectors were appended after conversion; the IQ2-only file cannot see images.
iq2_xxs, q6_K, q5_K, q8_0 and iq4_nl are the canonical GGML encodings. The DSpark experts keep the release's MXFP4 (fp4_e2m1 with e8m0 scales).
Vision path
Images enter through image content parts on /v1/chat/completions, the same OpenAI-compatible API as text. The chat template is in chat_template.jinja; the container must be served with --override-chat-template /models/chat_template.jinja.
The encoder is the release's own 32-block ViT and 3×3 aligner, kept bf16. Images are prefill-only and capped at 1,024 image tokens per image. The text path is the same IQ2_XXS model.
Serving
Hardware: four AMD Radeon AI PRO R9700 (gfx1201) cards, 32 GiB each; ROCm 10.1+; 64 GiB host RAM; the 259 GiB container and its 103 GiB Engram table on NVMe. The exact command, including --host-pool-mib 57344, --ngram-placement disk, --num-speculative-tokens 5 and --kv-cache-dtype fp8, is in RUNNING.md.
The server speaks OpenAI's chat API (/v1/chat/completions, /v1/models); the served model name is DeepSeek-V4.1-Flash.
Other radiance containers
- DeepSeek-V4.1-Flash-1017 uncensored UD-IQ1_S, 251.13 GiB — text-only, same abliteration, higher-width routed-expert mix.
Index: DeepSeek-V4.1-Flash radiance
Licence
DeepSeek-V4.1-Flash-1017 is MIT licensed (Copyright © DeepSeek). This container is a modified copy: the routed experts and linears are re-encoded as the quantisers above, the Engram table is generated from the tokenizer, the vision encoder is included at bf16, and 18 attention-output projections carry the abliteration edit. MIT applies to the original and to these modifications.