FYI: DeepSeek v4.1 Flash weights are available

[fry69]

Hugging Face link: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash

Tech report link: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/DeepSeek_V41_Tech_Report.pdf

config.json link: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/config.json

It seems those weights include vision by default now and it is a new "DeepseekV41ForCausalLM" architecture, including new model type and weights parts type slugs.

[pminervini]

it's enc/dec, adapting this codebase may be not so trivial

UPDATE: the man is already cooking :) https://x.com/antirez/status/2098121665771110540

[gilbert-barajas]

Regime: MacBook Pro M5 Max 128 GB (macOS 27.0), ds4 @ bd66c40, Q2 341 GiB
(SHA-verified), --metal --ssd-streaming --power 100; ds4-bench, promessi_sposi
sweep 2K→32K step 2K, 128 greedy tokens, cold page cache (first touch).
Prefill t/s: 72 / 47 / 56 / 68 — steady decode t/s: 15.8 / 15.7 / 15.7 / 14.8
at 2K / 8K / 16K / 32K. Warm 2K/4K points: prefill 130 / 112, decode 16.4 / 16.4.
Decode is flat vs context (14.8–16.4 t/s across 2K→32K): per-token cost in this
regime is streaming-dominated, and lands ~2.5× below resident V4 Flash on the
same box and sweep (34.9–40.0 t/s). Native tool_calls and 4-image vision both
pass through the OpenAI-compatible server (thinking on by default).
Cold-expert caveat: first-touch prefill collapses ~9× (112 → 12 t/s) and decode
decays 14 → 6.5 t/s as generation routes to never-cached experts — cold and
warm runs need separate labels on this build.