DeepSeek V4.1 Flash 1017 Support for ds4 · PR #1225 · antirez/ds4
Just pushed a PR for dwarfstar that adds support for Deepseek-v4.1-Flash-1017 with its new engrams and updated vision tower.
Proof1 community thread
Key points
- The PR basically just adds compatibility for the file revisions and checks to make sure the vision+language model file combination matches, raising on mixed pairings.
- Any files quantized with the ds4 quantization suite will load without any special changes, and it introduces no regressions for the previous checkpoint.
- This is a major revision, with ~40% of the routed expert tensors changed, 30% in the engrams, and ~50% of the vision tower's expert tensors changed.
- Shortly, I will be uploading my Q2-Q8 mixed quant for 128GB w/ minor SSD offloading (152GB wired, 366GB total with engrams, following antirez's recipe for the previous checkpoint), and will make a thread when it's ready.
Sources (1)
- [1]DeepSeek V4.1 Flash 1017 Support for ds4 · PR #1225 · antirez/ds4r/LocalLLaMA (top, daily) · Oct 17, 08:42 PM
Just pushed a PR for dwarfstar that adds support for Deepseek-v4.1-Flash-1017 with its new engrams and updated vision tower.
The PR basically just adds compatibility for the file revisions and checks to make sure the vision+language model file combination matches, raising on mixed pairings.
Extractive summary: sentences quoted from the sources.
Hype check
Independently evaluated4 independent looks since
- Evaluation · Oct 17, 2026Alignment on SOTA Chinese Models like Qwen 3.8 Flash Next / GLM 5.3 Flash
- Evaluation · Oct 17, 2026DeepSeek V4.1 Flash 1017 vs V4.1 Flash, Pi vs DeepSeek Harness: results on my 10 coding tasks
- Evaluation · Oct 17, 2026What is wrong with AI writing - ICLR edition [D]
- Evaluation · Oct 18, 2026T2I Comparison (~4MP): Qwen Image 2.1 Turbo vs. MOSS-Image-SFT vs. Krea 2 Turbo vs. Flux 3 Klein (🤮)
- Follow-up · Oct 17, 202632GB to 64GB of RAM took Qwen3.8-Flash-Next from IQ2_XS (2-bit) to Unsloth UD-IQ4_XS (4-bit) on a single 4090
- Follow-up · Oct 17, 2026Do we really need more 5B general-purpose image models?
Before this
- Oct 17, 2026deepseek-ai/DeepSeek-V4.1-Flash-1017
- Oct 17, 2026Nemotron 3.5 Bolt, Muse Spark Open, and a Fast LFM2.5 24B
- Oct 16, 2026Qwen Flash Next potential SOTA Q2 nobody talks about
- Oct 15, 2026QUASAR 4-bit QAT checkpoints for Qwen3.5, Gemma 4 and Muse Glimmer: closer to BF16 than Google’s and NVIDIA’s 4-bit releases
- Oct 15, 2026nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB · Hugging Face
- Oct 15, 2026NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning
