ds4 PR adds DeepSeek V4.1 Flash 1017 support: ~40% routed experts changed
returnity · reddit · 2026-10-18
Developer returnity pushed a PR to antirez's ds4 adding support for DeepSeek-v4.1-Flash-1017 with its new engrams and updated vision tower. It's a major revision: 40% of routed expert tensors, 30% of engrams, and 50% of the vision tower's expert tensors changed, though the architecture is unchanged so existing kernels work as-is. The PR validates vision+language file pairings and loads ds4-quantized files without changes. The author plans a Q2-Q8 mixed quant for 128GB RAM machines (152GB wired, 366GB total), estimating 20 TPS decode on an M5 Max. Also a credible signal that the new DeepSeek checkpoint is real.
More from Infra
- Rust-based Git server on S3 handles 7,600+ concurrent pushes per second — blaizedsouza · 2026-10-18
- Chaining 2 PCs, a MacBook and 2 iPhones Runs 48-Expert MoE at 37 tok/s on 60k Context — StayLameBro · 2026-10-18
- Ollama US hosting coming soon, bringing DeepSeek-V4.1-Flash-1017 stateside — gaganghotra_ · 2026-10-18
- 43-minute video encoded to AV1 in 2m35s on Intel Arc A380, 77% smaller — haydendevs · 2026-10-18
- NYT: CoreWeave's financial arrangements likened to WeWork as AI debt worries grow — SumitGup · 2026-10-18
- 32GB→64GB RAM lets a single RTX 4090 run Qwen 4-bit quants at 90+ tok/s — Distinct-Pie2389 · 2026-10-18