DeepSeek V4.1 Flash hands-on tests and comparisons

A deepseek-v4.1-flash call returns a new system_fingerprint matching the 1017 build, with routing_info and X-Ds-Trace-Id showing direct official API passthrough, so client model IDs stay unchanged.

A post says V4.1 Flash is free with unlimited usage in the OpenCode desktop app on macOS and Windows, with no subscription or API cost.

Experts are losslessly transcoded from MXFP4 to NVFP4 at about 430GB, fitting on four DGX Sparks with roughly 80GB left for KV cache, with a two-Spark build reportedly in the works.

Running the classic pelican-on-a-bicycle prompt through Codex and DeepSeek Harness produced an animated SVG turned into an 11-second video, using ~2.87M tokens at 92% cache hit rate for roughly one yuan.

Release roundup: upgraded one-sentence-to-full-project coding, web/app/API updated at the same price with open weights; the author plans to test it hands-on.

Official release: MIT-licensed open weights, with vendor-run evals showing Terminal-Bench 4.0 rising from 31.2 to 42.3 and Apex Bench from 64.6 to 72.1.

The poster says the model is live on DeepSeek's platform and open-sourced on Hugging Face, linking pricing docs and calling the test results impressive.

A reply to SipeedLab points to a GitHub TP4 recipe as the strongest four-machine ds-4.1-flash setup, suggests sparkdash screenshots for a common benchmark, and says the Chinese community widely calls the claims inflated.

A four-point take on the upgrade: shorter outputs cut inference cost, tooling optimizations target DeepSeek Harness and CodeBuddy, strong DeepSWE and Terminal-Bench scores reinforce the price-performance pitch, and same-day rollout across site, app and API drives usage.

The user argues DeepSWE has become a joke of a benchmark and manually inserted V4.1 Flash 1017 results into a CodingFleet blog screenshot.

A user says after a Claude ban they used DeepSeek V4.1 Flash with open-source OpenDesign Motion to produce a brand motion video from one prompt, calling it an Opus 5.5 alternative.

A long Japanese post analyzes performance, pricing, the 763B open weights and post-training, citing 77% DeepSWE success versus about 75% for GPT-6.1 Sol at roughly one-third the cost.