DeepSeek releases multimodal model with vision agent capabilities rivaling Opus-4.8
赛博禅心 · wechat · 2026-08-21
The DeepSeek multimodal model (deepseek-v4-flash-vision-expbench) is now available via API and DeepSeekHarness, maintaining the same pricing as ds4f.
Benchmark Scores:
- TerminalBench 2.1: 83.9
- NL2Repo: 57.7
- DeepSWE: 59.3
- DSBench-Hard: 63.6
- AutomationBench (Public): 25.7
- ApexBench (Pass@1): 36.5
- Agents' Last Exam: 27.3
- Chartography: 64.3 (p0.95); 63.3 (p1.0)
- ZeroBench (Pass@5): 35.0
The new model shows significant improvements in vision-based Agent benchmarks compared to ds4f, closing in on the capabilities of Opus-4.8.
Related event: DeepSeek Launches V4-Flash-Vision-Exp, Closing In on Opus 4.8(29 posts)→
More from coding & agent
- Open source MCP server implements HTTP 402 micropayment gateway — EstablishmentTough18 · 2026-08-23
- nb.nvim Fixes Cross-Folder Image Link Bug — 4310sy · 2026-08-23
- AI Agent Search Optimization: Path Calculation Drops from 504ms to Near 0ms — DanielLockyer · 2026-08-23
- How To Build An AI Agent: A Complete Workflow Guide — mdancho84 · 2026-08-23
- A practical guide to AI agents for small businesses: lead response in minutes, 40 hours of invoice work down to 6 — juliesweetie93 · 2026-08-23
- GSD Pi: A terminal agent for long-haul autonomous software projects — tom_doerr · 2026-08-23