DeepSeek V4.1 Flash hands-on tests and comparisons

The post lists V4.1 Flash 1017 benchmarks and specs: 77 on DeepSWE beating Opus 5, 42.3 on Terminal Bench still behind, 1M context, MIT-licensed free weights, live on app and web.

The author is waiting for DeepSeek V4.1 Flash 1017 on Ollama Cloud and currently maintains open-source projects via the official DeepSeek API, with no concrete data.

A user notes that V4.1 Flash 1017 uses sparse attention attending only to top-k tokens per block instead of full context, cutting compute while keeping coherence.

Versus the previous version, coding, terminal use, agent execution and complex tasks improved notably, beating GPT-5.6-Sol on several specialized benchmarks, though Claude Opus 5 still leads on general reasoning and hard coding.

Claims DeepSeek V4.1-Flash surpasses Claude Opus 5 on coding at about $0.2 per task versus over $10 for Claude, releasing 763 billion parameters for free under the MIT license.

The post says the release improves 3D, coding, agents and vision, ships faster and shorter outputs at the same price with 1M context, is live on app, web and API, and has MIT-licensed open weights.

Official figures show Terminal-Bench 4.0 at 42.3 vs 31.2 and DeepSWE v1.1 at 77.0, beating Opus 5's 74.0; the API now serves deepseek-flash.

DeepSWE 77% vs Opus 5's 74%, demos include a Forbidden City 3D scene and a Godot Portal 2 remake; Search/Fetch API bills 5 yuan per 1,000 successful calls, but users say it follows instructions literally and won't add UI/UX polish.

A user reports DeepSeek V4.1 Flash 1017 scoring 77% on DeepSWE at a much lower price than closed models, citing open weights and agentic gains.

Official release: 763B params, 1M context, MIT open weights; DeepSWE 77.0 vs Opus 5's 74.0 at about $0.2 per task versus $10+, though it still trails on HLE and Terminal-Bench.

A rundown of the official release: MIT-licensed weights, native image-text FP8, 1M context and 256K output, DeepSWE 77 beating Opus 5 while trailing on Terminal-Bench and HLE, plus Search & Fetch at $0.75 per 1,000 calls and free in Harness, Claude Code and Codex until November 1.

The author argues self-feedback reinforcement learning lets V4.1 Flash 1017 build complete projects from scratch in game dev and 3D modeling, enabling wish-style coding.