Day 7 receipt
DeepSeek-V4.1-Flash-1017
Day 6 of 7DeepSeek's update to V4.1 Flash, released as MIT-licensed open weights and as the model behind its website, app and API.
Launch dayWhat they claimed, quoted word for word · 4
“Now live on our website, app, and API.”
Testers report using it through the API inside an agent harness, and the open weights are public and ungated on Hugging Face.
“Stronger benchmark performance, faster task completion, and shorter outputs than the previous version”
One independent run found shorter outputs at every reasoning level with the same quality. The benchmark and speed parts are not yet checked by users.
“DeepSeek-V4.1-Flash-1017 uses reinforcement learning with self-feedback to improve end-to-end project development”
Early coding reports are mixed: one tester had it outscore two Claude models, another got a better result from a Qwen model. Too few to call.
“This repository and the model weights are licensed under the MIT License.”
The weights are downloadable without an access form, and one tester ran them locally. The release has no chat template file, so users must reuse the earlier one or DeepSeek's own encoder.
Day 7What real users report · 7
A tester who benchmarked the new weights against the old ones found fewer output tokens at every reasoning level, with the lowest level now usable, and the same intelligence and tool-calling pattern.
Comment in a 953-point, 716-comment thread · Open the source
A user ran three hard questions with it and two Claude models as worker agents in Claude Code; it came out cheaper and more correct than the mid-tier Claude model.
Comment in a 953-point, 716-comment thread · Open the source
A user who ran it over a weekend for simulation code found it good, but got a better result from a Qwen model on the same task, and called it too early to judge.
Comment in a 186-point thread · Open the source
A user ran the open weights locally on a consumer GPU by streaming from SSD, but said it was too slow for them to judge quality.
Comment in a 186-point thread · Open the source
A user could not find a chat template in the repository and asked whether to reuse the one from the previous Flash release; a reply said yes.
Hugging Face discussion · 2 comments · Open the source
A user who found the original V4.1 Flash worse than GLM 5.3 on their own codebases said this build, after 48 hours on dozens of tickets, beat GLM 5.3 and was faster on the same four-GPU hardware.
Hugging Face discussion · 1 comment · Open the source
A user published a patch for SGLang v0.5.22 to serve the model on four RTX Pro 6000 GPUs with a 524,288-token context; the only follow-up asked for speed numbers.
Hugging Face discussion · 2 comments · Open the source