DeepSeek V4.1 Flash hands-on tests and comparisons

The post points out the new model's 763B total / 196B Engram / A16B active size versus V4 Flash's 284B/A13B, linking the Hugging Face weights release.

TensorX says it deployed V4.1 Flash 1017 on its own B300 GPUs in Europe the same day, citing a DeepSWE v1.1 score of 77.0, above Claude Opus 5.

A user spent a morning with V4.1 Flash producing a video with edits, animations, cover image and audio compression for $0.30, judging its skills on par with Astra and Fable but the model slow.

A user reports DeepSeek V4.1 Flash 1017 hitting 77.0 on DeepSWE versus 66.9 for GLM-5.3, right after GLM-5.3 HighSpeed entered gray testing.

Official release: the new model is live on web, app and API with higher benchmarks, faster task completion and shorter outputs than the previous version.

A usage take: let Opus 5.5 handle reasoning and planning while DeepSeek V4.1 Flash 1017 does the execution, which the user calls sufficient.

A roundup of day-one free access: Experiential Labs offers V4.1 Flash 1017 free over API with 1M context, and OpenCode Zen lists it at $0 for a limited time, though Zen's free lane rejects other clients.

A user who tried DeepSeek-v4.1-flash-1017 in DSH reports it is quite slow, even in low mode.

A developer's first test runs the EXL3 2.9bpw quantized build on two DGX Sparks, with a screenshot but no throughput numbers.

The author ran all 157 tests through a Hermes agent with thinking off: 128 points, one better than September, but math fell to 12 of 30, and only effort none actually disables thinking.

A tester reports V4.1 Flash 1017 executed a task autonomously for 31 minutes on one prompt for $0.22, producing a Three.js SpongeBob scene with water absorption, physics and cinematic lighting.

Asked to code a 30-second video of frontier releases since 2025, the model spent 3 hours and produced a result the author found disappointing.