DeepSeek V4.1 Flash hands-on tests and comparisons

Specs: 1M-token context, up to 393K output tokens, text and image input, four reasoning effort levels, and $0.15/$0.60 per million input/output tokens, unchanged from September, with no official announcement yet.

Weights are on Hugging Face under MIT; the official table shows DeepSWE 77.0, Automation-Bench 60.5 and Agents' Last Exam 35.4 above Opus 5, while HLE is 42.6 vs 56.3 and Terminal-Bench 42.3 vs 51.8; self-hosting needs encoding.py for prompts.

The post says the new version beats K3, GLM-5.3 and GPT-5.6 Sol on reasoning, coding, terminal agents and multimodal vision, with shorter faster outputs, a deepseek-recipe library and continuous reasoning_effort 1–100.

On OpenRouter, V4.1 Flash 1017 costs $0.15 input, $0.60 output and $0.003 cache read per 1M tokens, which the poster calls Kimi K3/GLM 5.3-level quality at a bargain price.

A new V4.1 Flash checkpoint has landed on DeepSeek's official Hugging Face page, following the previous day's API leak.

A user notes the DeepSeek-V4.1-Flash-1017 weights are on Hugging Face at 510GB with no GGUF quantization available yet, complicating local deployment.

The official post links Hugging Face and ModelScope downloads, confirming open weights for DeepSeek V4.1 Flash-1017.

On Hugging Face under MIT, it beats Opus 5 on DeepSWE 77 vs 74 but trails on Terminal-Bench 4.0 at 42.3 vs 51.8 — a big jump, not yet Opus level.

A user inspecting the Hugging Face weights estimates 510GB total with 204GB in Engram, targeting around 30 tok/s if trimmed to 306GB.