DeepSeek-V4.1-Flash tops coding benchmarks: 28.5% on FrontierCode v2.0, beats Kimi K3
soumithchintala · x · 2026-11-06
PyTorch creator Soumith Chintala shared a coding/autoresearch benchmark comparison, noting the model also works for hours autonomously on GPU kernel optimization.
Key results across five models:
- FrontierCode v2.0: DeepSeek-V4.1-Flash leads at 28.5%, vs Kimi K3 17.3%, GLM-5.3-Flash 14.2%, Muse Spark Open 7.6%
- DeepSWE v1.1: DeepSeek-V4.1-Flash 77.0%, Inkling 2 Small 72.1%
- SWEBench Pro v2 (Hard): Kimi K3 tops at 88.2%, Inkling 2 Small 86.3%
- Vibe Code Bench v1.1: DeepSeek-V4.1-Flash 87.7%
- Terminal Bench Science v0.1: DeepSeek-V4.1-Flash 12.9%, well ahead of all others (below 8.1%)
Overall DeepSeek-V4.1-Flash takes first place on most coding and agentic benchmarks, with a notable edge on terminal science tasks.
More from coding & agent
- Stripe Link now works with nearly 400 agents, auto-optimizing payment per transaction — jeff_weinstein · 2026-11-06
- Replit CEO: we're the only collaborative AI coding system, unlike Claude Code — amasad · 2026-11-06
- Consultant charges $20,000 to install a Claude 'Business Brain' covering 3 full-time roles — aryanXmahajan · 2026-11-06
- Spin up your own media server on Cloudflare for ~$6/mo — just ask your agent — PurzBeats · 2026-11-06
- Inkling 2 Small matches models 8-13x pricier at just $0.34 per task — echen · 2026-11-06
- Replit's custom kernel sandbox dodged 96% of 6,211 audited Linux CVEs — amasad · 2026-11-06