DeepSeek V4.1 Flash hands-on tests and comparisons

A user notes that Maya has begun supporting the 748B MoE DeepSeek-V4.1-Flash-1017 and looks forward to local use.

ivanfioravanti tuned ds4 on M3 Ultra: Q4 two-Mac prefill reaches 920–1,193 t/s and decode 48.5 t/s, with logprobs within noise of the official API across 84 cases.

Project Maya adds support for V4.1 Flash 1017: a ~500GB standard GGUF plus a 178GB Maya-S quant with 203GB Engram, with memory-allocation issues reported on Windows.

One-shot, harness-free Three.js anime-inspired diorama with lo-fi music and ambience under 100k tokens, matching the prompt's scene type exactly.

A Japanese AI news roundup says the 1017 build scores 77% success on the DeepSWE v1.1 coding-agent eval, beating GPT-6 Astra and GPT-6.1 Sol, with MIT-licensed weights and unchanged $0.15/$0.60 API pricing.

A Japanese post says DeepSeek V4.1 Flash 1017 is now free in OpenCode via Google sign-in, recommends using it to build a browser FPS game, and calls it stronger than most US models at coding.

On Agent Arena's real agent-task leaderboard, DeepSeek 1017 ranks 11th globally and first among open models, with net improvement rising from 2.56% to 5.13% and user-confirmed completion from 6.60% to 11.54%.

DeepSeek V4.1 Flash 1017 is now callable for free in OpenCode, though speed, context length and commercial-use terms are not yet stated.

The author spent 93.9M tokens of DeepSeek V4.1 Flash in DeepSeek Harness to rebuild the 90s castaway screensaver as an open-source GitHub project.

Running V4.1-Flash-1017 with the TensorFold engine and MIA AI Lab's EXL3 quants on two DGX Sparks gave 107 tok/s aggregate over 4 concurrent lanes, about 28 tok/s each, which the author calls very usable desktop-level serving.

A Base daily recap notes Venice upgraded its app and API to the 1017 build of DeepSeek V4.1 Flash, stressing full privacy and claiming it beats leading open and frontier models on most benchmarks.

The author generated a video directly through DeepSeek V4.1 Flash-1017 in Qoder during its free window and says Qoder's handling of the model's chain-of-thought looks oddly 'reverse-optimized'.