DeepSeek V4.1 Flash at 2.61 bpw + Engram: 26 t/s on dual RTX 3090s with dsx
Lucebox ROCMFP 2.5 bpw 0.253 78.7% 9.78
Proof1 community thread
Key points
- The model's weight has been uploaded.
- Author's 2.03 bpw build 2.03 bpw 0.462 71.5% 11.66
- The engine doesn't seem to support images yet.
- At 262,144 context with Engram tables (203 GB) on a SATA SSD, I got 26 t/s generation speed and 170 t/s prompt processing speed.
Sources (1)
- [1]DeepSeek V4.1 Flash at 2.61 bpw + Engram: 26 t/s on dual RTX 3090s with dsxr/LocalLLaMA (top, daily) · Oct 19, 11:43 AM
Lucebox ROCMFP 2.5 bpw 0.253 78.7% 9.78
The model's weight has been uploaded.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 17, 2026[AINews] High Taste Testers Ep. 1
- 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, 2026Welcome Beam: a 501B open model you can own, customize, and scale
- Oct 15, 2026Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint
- Oct 15, 2026NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning
