DeepSeek V4.1 Flash hands-on tests and comparisons

A breakdown of the release: self-feedback RL lets the model debug and fix its own bugs with fewer interaction rounds, while KV cache compressed to 890 bytes per token cuts prefill compute nearly in half, making 1M context cheaper.

Item-by-item: DeepSWE 77.0 versus 74.2 prior, Sol 73.0 and Opus 5 74.0; Terminal-Bench 4.0 42.3 versus 31.2 prior, Sol 39.9 and Opus 5 51.8; Automation-Bench 60.5 versus 54.8 prior, Sol 45.8 and Opus 5 50.3.

User recap: V4.1-Flash-1017 tops several agent coding benchmarks over Opus 5, costs about $0.2 per task, ships MIT-licensed weights, and bundles free Search and Fetch APIs until Nov 1.

Notes that 1017 is a post-trained update, not a new architecture: HLE 42.6 (from 36.8), Terminal-Bench 42.3 (from 31.2), DeepSWE 77.0 (from 74.2), same deepseek-flash API at $0.003 cache hit, $0.15 miss, $0.60 output.

OpenCode says DeepSeek V4.1 Flash 1017 is now free to use in its coding tool, with no further details given.

A Japanese hands-on reports 99% cache hit on the first turn, over 200 tok/s, and that the model can be jailbroken.

The poster says the 552B MoE V4.1-flash-1017 beats Opus 5 on DeepSWE and Automation-Bench while costing $1.20/M output tokens ($0.60 off-peak), over 20x cheaper.

An English post relays the official update: stronger coding, 3D, visual understanding and agent workflows, shorter and faster outputs, and a 1M-token context at unchanged API pricing, though the author has not tested it yet.

Terminal-Bench 4.0 rose 31.2 to 42.3, ProgramBench 20.3 to 29.1 and DeepSWE v1.1 74.2 to 77.0, above GPT-5.6 Sol's 73.0 and Opus 5's 74.0, with shorter task time and output; the author cautions against calling it a total sweep.

The poster argues 1017 is a brand-new version rather than a minor update, comparing its wish-granting coding feel to Opus 5.5 and Fable 5.1.

Official charts show Terminal-Bench 4.0 rising from 31.2% to 42.3% and DeepSWE at 77.0%, with unchanged API pricing, shorter outputs and MIT open weights already live.

Official charts put it at 77.0% on DeepSWE v1.1, above Opus 5's 74.0 and GPT-5.6-Sol's 73.0, while still trailing Opus 5 on HLE and ProgramBench; top tier costs about $0.2 per task.