DeepSeek V4.1 Flash hands-on tests and comparisons

The author upgraded their Martin cluster using a GitHub vLLM DGX Spark recipe for DeepSeek-V4.1-Flash and posted a bugswat report on the local deployment.

The author claims 77% DeepSWE success at about $0.20 per task versus roughly 75% and $0.65 for GPT-6.1 Sol, input at $0.15 and output at $0.3 per million tokens, and 763B weights (~510GB) runnable on four DGX Sparks or a 512GB Mac Studio.

Official numbers show wins on four agentic benchmarks but losses on HLE, Terminal-Bench, ProgramBench, NL2Repo and vision; the real edge is MIT open weights at $0.30/$1.20 per million tokens.

A French post relays the release: V4.1-Flash-1017 improves coding, reasoning and tool use, with MIT-licensed weights on Hugging Face.

Three models each built a gummy bear webpage from the same reference, compared side by side on stretch, throw and collision behavior, with viewers asked to vote.

Using a direct API call with no harness, external assets or libraries, the model generated a retro-style game entirely in code in one HTML file, which the author calls a huge quality upgrade.

A reusable motion-design prompt that turns DeepSeek V4.1 Flash 1017 specs — 552B MoE, 1M context, DeepSWE 77.0, $0.20 per coding task — into a beat-synced product reveal video.

A summary of the official release: DeepSWE 77.0 (from 74.2) and Terminal-Bench 42.3 (from 31.2), MIT open weights, 1M context, API at $0.15/$0.60 per 1M tokens off-peak (2x at peak), and Search/Fetch at $0.75 per 1k calls.

Same rocket launch scene test: DeepSeek cost $0.78 and 3 minutes versus GPT 6.1 Sol's $6.10 and 5 minutes, nearly 8x cheaper and faster.

An Experiential Labs model page advertises free limited-time access to V4.1 Flash 1017 with a 1.05M-token context window and up to 393K output tokens.

Single-stream benchmark deltas for the 1017 build: code 73.8→76.3, format 55.4→75.8, JSON 52.1→63.2, reasoning 37.8→55.6, with a vLLM DGX Spark repo.

A roundup of models shipped in the first 17 days of October, listing Gemini 4 Argon alongside Claude Haiku 5.5 and the Muse lineup, plus several still pending.