DeepSeek V4.1 Flash hands-on tests and comparisons

A harness-free one-shot Three.js campfire using no external assets and 81k output tokens, with the embers singled out as especially good.

For models larger than local hardware supports, the post recommends project-maya for DeepSeek V4.1 Flash 1017 or GLM, and strata for Qwen 3.8 Flash Next.

bradthomasbrown says DeepSeek V4.1 Flash 1017 outperforms GPT-6 Astra on an ML research test where an agent builds a model to solve a benchmark instead of solving it directly.

On dual 3090s, the dsx 2-bit build hits 37 tok/s while Unsloth's 1-bit GGUF gets 12.3 tok/s; the author blames lossy quantization for thinking length rising from 10.7K to over 20K tokens.

A user running DeepSeek V4.1 Flash 1017 across TP4 with a ring mesh reports 100 tok/s without switches and asks how to go faster without losing quality.

A quantized build for Project Maya keeps hot experts on GPU and streams the rest from RAM and SSD; the author measures 92.6% next-token agreement with the original (KL 0.120) under an MIT license.

Artificial Analysis reports raw red-team conversation scores from 30% for DeepSeek V4.1 Flash (max) to 65% for GPT-6 Astra (max), versus 63% to 79% on good-faith conversations.

On 90 hard BFCL tool-call, Omni-MATH 7+ and SuperGPQA items graded by code, Beam scored 67/90 and DeepSeek 66/90, though Beam costs about 6x more.

Project Maya added DeepSeek V4.1 Flash and measured 2,000 tps prefill and 50 tps decode at 256k context, slightly slower than GLM-5.3-Flash on the same hardware, with room for optimization.

Asked to write a song about itself and animate it in JavaScript, V4.1 Flash 1017 used 20M tokens and cost $0.30 to produce the result.

A user says DeepSeek-V4.1-Flash has been merely average so far and feels a level below GPT-6 Sol with daybreak, without giving specific test details.

Databricks says DeepSeek V4.1 Flash 1017 is available as a new endpoint with updated weights and gains in reasoning, coding, agentic and visual benchmarks.