DeepSeek V4.1 Flash hands-on tests and comparisons

A user runs the Maya-XS quant of DeepSeek-V4.1-Flash-1017 via Project-Maya on an RTX 3090 24GB plus RTX 5070 Ti 16GB with 96GB RAM, marveling at local LLM progress.

With oMLX 0.7.2, OQ4 quantization plus MTP let an M3 Ultra run a 1M-token context on DeepSeek V4.1 Flash, which the author calls practical for local work.

The New Stack reports DeepSeek V4.1-Flash-1017 was more accurate than Claude Haiku 5.5 in its test while costing a fraction of the price.

A GCP deployment script runs the abliterated DeepSeek V4.1 Flash in FP8 on four RTX PRO 6000 GPUs at roughly 200 tok/s.

On Cognition's FrontierCode v1.1 benchmark of 100 real open-source software engineering tasks with a private test set, DeepSeek V4.1 Flash 1017 MAX ranks above GPT-6.1 Sol MAX.

Testing local LLMs inside Maya, the author ranks DeepSeek-V4.1-Flash IQ2_XXS at least equal to GLM-5.3-Flash IQ2_XXS and above V4.1-Flash IQ1_S, and plans to use it in real work.

After heavy use, the poster finds DeepSeek Flash 1017 can't reliably finish long-horizon tasks alone, but pairing it with Opus or Astra as advisor/commander greatly improves completion at good cost efficiency.

In a shared 3D beach diorama build, Opus 5.5 ranked first at ~1h50m and ~$29, GPT-6.1 Sol second at ~2h51m and ~$5.35, while DeepSeek V4.1 Flash 1017 ran 5 parallel agents for 4h19m and 124M tokens at just $2.47, with the author saying a few more fix prompts could close the gap.

A Japanese post notes the API name stayed the same while the model behind it changed, laying out DeepSeek V4.1 Flash 1017's changes and what to check via success rate, review time, and total task cost.

The author shares an updated coding benchmark screenshot showing DeepSeek 1017 Flash in second place behind Opus 5.5 and ahead of GPT-6.1 Sol and Astra, noting it ran in the Codex harness and would do better in Pi or DSH.

The author tested DeepSeek-V4.1 Flash 1017 locally on oMLX 0.7.2 with OQ4 quantization and MTP, finding it worth switching to but noting weaker critique on fiction writing.

The author notes Claude Haiku 5.5 input at $0.1/M is cheaper than DeepSeek V4.1 Flash's official off-peak price, yet DeepSeek still tops OpenRouter token usage, suggesting price is no longer its moat.