DeepSeek V4.1-Flash-1017 breakdown: 763B open weights, coding cost down ~60%, search API opens
赛博禅心 · wechat · 2026-10-17
A breakdown of DeepSeek's V4.1-Flash-1017 update: 763B parameters, MIT-licensed open weights, interleaved image-text input, 1M context (recommended temperature 1.0, topp 1.0, maxtokens ≥256K), rolled out across web/app/API at unchanged prices.
Performance & cost: gains across all 11 benchmarks; DeepSWE v1.1 max tier passes 77% at 1.3 RMB/task (vs 74.2% at 3.3 RMB for the old version; low tier 72.6% under 0.8 RMB), while Claude Opus 5 max costs 80 RMB per task. Optimized for DeepSeek Harness (upgrade to 0.2.1-rc.1+; 6 RMB credit for new logins), Tencent CodeBuddy and WorkBuddy.
Wish-style coding: self-feedback RL for from-scratch projects — racing game demo, Godot Portal 2 recreation, Valley of Ages evolution scene.
Search API: Search/Fetch at 5 RMB/1k calls, filtering SEO spam and GEO manipulation; server-side calls in Harness/Claude Code/Codex become billed Nov 1.
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