DeepSeek V4.1 Flash hands-on tests and comparisons

A user compiled output speeds for running DeepSeek V4.1 Flash across different implementations on a Mac Studio M3 Ultra 512GB, crediting Unsloth for making the model usable.

On 27 real assistant tasks with three trials each, both cost about 1.3 cents per task; Beam's p90 turn was 22s versus DeepSeek's 87s, and DeepSeek's lead came mainly from admitting missing access.

OpenCode is offering 14 models free for a limited time; DeepSeek V4.1 Flash and some others are stated not to retain or train on data, while MiMo and others may use free-period data for model improvement, so sensitive code should be avoided.

Ivan Fioravanti reports DeepSeek V4.1 implementing a C64 game zero-shot, stressing that pairing it with DeepSeek Harness makes a big difference, and says he tried the same prompt locally on an M3 Ultra 512GB and with MiniMax M3.1.

Project Maya's maya-ds4 engine ships an 8GB dspark drafter (fp4 experts, q8 dense) for the 340GB DeepSeek-V4.1-Flash-1017 GGUF, with 92.6% top-1 match, and only drafts when it beats base decode.

A 1.62-bit quant of DeepSeek V4.1 Flash ran fully locally on one ASUS Ascent GX10 to build and refine four small games, with gameplay recorded.

iFlytek released Spark X2.6 (293B-A30B, 1M context) on domestic compute and open-sourced a 4B edge model, claiming code and agent parity with DeepSeek-V4.1-Flash, though no third-party benchmarks exist yet.

A user reports DeepSeek V4.1 Flash 1017 running locally at 675.9 tokens per second, a speed they say most people still consider impossible.

Official docs confirm $0.15 input and $0.60 output per million tokens off-peak, doubling during weekday JST 10–13 and 15–19, so the author schedules non-urgent test-code generation into cheaper hours.

A user relays the release: same Flash pricing, claimed better scores and shorter runtimes, plus an open Search & Fetch API, while questioning whether rate limits stay stable.

A user ran DeepSeek-V4.1-Flash-1017 via deepseek-api at high thinking effort with no harness, producing six rich animated forest scenes in single one-turn runs of roughly 100–130k tokens each.

A user finds a reseller charging $0.144 input and $0.576 output per million tokens for V4.1 Flash 1017, about 48% of official pricing and effectively 24% with top-up bonuses, and calls its capability close to Opus 5.5.