DeepSeek V4.1 Flash hands-on tests and comparisons

All kernels rewritten in TileLang with bit-identical outputs; prefill up 12–14% on four Sparks, 75 tok/s single stream, two logprob bugs fixed, 5,730 tests passing.

On a Python Doom-style game rendered in tmux Braille, DeepSeek V4.1 Flash 1017 at ~40 tok/s on 2×3090 kept producing results while Nanbeige4.3-33B at 70 tok/s on one 5090 overthought.

A provider-by-provider price check finds Cloudflare never pricier than official, DeepInfra fp8 cheaper but capped at 131K output, most others at peak pricing, plus cache-price and routing traps with whitelist and max_price configs.

A quantizer reports their EXL3 quants took the top three spots on a testing site, with a 1.62-bit DeepSeek V4.1 Flash EXL3 beating their lossless Qwen Flash Next EXL3.

TensorFold runs the full official weights on a switchless 4x DGX Spark TP4 setup with 1M context and 16 concurrent sessions, reporting per-concurrency decode speeds and 4000+ tok/s prefill.

Two open-source local runners updated within hours: Strata decodes about 10-20% faster when the model fits in RAM, while Maya reports up to 2x on an RTX 3090 and 14% on a V100, Linux single-GPU only.

Testing DeepSeek-v4.1-flash-1017 on destructible game assets and explosion effects in three.js, the author got everything from a single prompt in one text output with no harness or tool calls.

A user notes DeepSeek V4.1 Flash is trending for high medical benchmark scores at very low cost with MIT-licensed open weights, says the 1017 checkpoint surprised them in testing, but still picks Sol for quality-critical work.

The author describes a voice-prompt plus visual-reference app building loop, mixing Claude Code with Opus 5.5 and OpenCode with DeepSeek V4.1 Flash / GLM 5.3 as harnesses.

The author had both models build in his own voxel engine from the same prompt, each in 10 seconds, with half cut away to compare spatial reasoning.

A user says V4.1 Flash 1017 may be S-tier, praising its reasoning style, fast 3D/visual performance and $0.10 price.

The author says the Maya models passed 410k HF downloads and, after enabling fast local GLM-5.3-flash runs on small devices, now support DeepSeek-v4.1-flash as well.