DeepSeek V4.1 Flash hands-on tests and comparisons

Scored on features, visuals and code quality, V4.1-Flash-1017 and GPT-6 Astra both land near 7.5 with Opus 5.5 at 7.4; eight DeepSeek tests cost about $0.71 versus an estimated $31.69 for Astra.

A user says DeepSeek V4.1 Flash shows as free in WorkBuddy international and can be reverse-proxied with magpie for use in other agents.

The report contrasts V4.1-Flash-1017's 552B sparse MoE, ~890 bytes/token KV and low API rates with KUT Gemini Flash's Ricci Flow manifold constraints, weighing per-token cost against formal-verification reliability.

Citing official charts, the post notes DeepSWE 77 versus Opus 5's 74, about $0.20 per coding task, and off-peak output at $0.60 per 1M tokens against Opus 5's $25.

A rundown of the official release: open MIT weights, 1M context, DeepSWE v1.1 at 77.0 versus Opus 5's 74.0, Terminal-Bench 4.0 up from 31.2 to 42.3, still behind Opus 5 on HLE, and Search/Fetch at $0.75 per 1,000 calls.

Official numbers show Terminal-Bench up from 31.2 to 42.3 and ProgramBench from 20.3 to 29.1, with shorter outputs, same API pricing, and web/app already updated.

A local deployment video runs the ~748B-with-Engram model on two asymmetric RTX 3090s (48GB VRAM) and 128GB DDR4 at about 30 tok/s, calling it a Sol-level model on an old gaming PC.

A user says the interface showed DeepSeek V4.1 Flash-1017 but the model claimed to be an Anthropic Claude model, and it outperformed GPT-6.1 Sol and Sonnet 5.5 on some tasks, raising questions about backend routing.

The post cites vendor-reported DeepSWE v1.1 scores of 77.0 for the open-weight model versus 74.0 for Claude Opus 5 and 73.0 for GPT 5.6-Sol, urging self-testing first.

On the same abandoned-village cinematic video task in pure JS, V4.1 Flash cost $2.5 in 4 hours versus Opus 5.5's $54.2 in 7 hours.

The author notes MIT weights, shorter outputs and unchanged pricing, and asks if anyone has tested it on real repos in Claude Code or OpenCode.

A Korean post summarizes the launch: web, app and API availability, MIT open weights, DeepSWE 77.0 versus GPT 5.6 Sol's 73.0 and Opus 5's 74.0, and under $0.20 per task, noting the numbers are self-reported.