DeepSeek ships V4.1-Flash-1017: 77% on DeepSWE beats GPT-6 Astra at ~$0.2 per task
新智元 · wechat · 2026-10-17
DeepSeek quietly released V4.1-Flash-1017 across web, app and API at unchanged prices, open-sourcing the 763B weights (MIT license, 1M context, interleaved image-text input) on HuggingFace and ModelScope.
Headline upgrade: "wish-style" coding
- Trained with self-feedback RL to build complete projects from scratch: demos include a Forbidden City 3D scene, a browser racing game, a Godot-based Portal 2 recreation, and a 3D timeline from prehistory to modern city
- DeepSWE v1.1: 77%, beating GPT-6 Astra (74%) and Claude Opus 5, at roughly $0.2 per task — what DeepSeek calls a "Pareto improvement" on performance-cost
- Terminal-Bench 4.0 rose from 31.2 to 42.3, with shorter runtimes and outputs
Self-built search API goes public: indexes hundreds of billions of pages, filters SEO spam and low-quality AIGC. Search and Fetch each cost 5 RMB/1k calls; server-side calls inside DeepSeek Harness, Claude Code and Codex stay free until Nov 1, 2026.
More from coding & agent
- Palo Alto Networks ships Cortex XCOR: AI SRE agent diagnoses alerts in under 3 minutes — shashib · 2026-10-17
- HF dev begs OpenAI/Anthropic to fix agents picking the wrong Chrome window — NielsRogge · 2026-10-17
- An agent negotiated a 3PL contract in 3 days, end to end — aarthir · 2026-10-17
- Gemini 3 vs Opus 4.5 vs GPT-5.1 vs Grok 4: same-prompt 3D diorama showdown — cedric_chee · 2026-10-17
- Reverse-engineering OpenAI Atlas browser's system prompt with Claude Code — DionysianAgent · 2026-10-17
- Hugging Face Hub adds RL Environment category with 153 agent-ready environments — mervenoyann · 2026-10-17