JevBench v1.8 crowns mammoth-1 atop open-weights board, now prices models by GPU time
airesearch12 · x · 2026-10-30
JevBench v1.8 ships two key changes:
- New open-weights leader: baseten's mammoth-1 tops the Jev-class capability board at 85.2 (Intelligence 76.2 / Calibration 57.2), at roughly $0.027 per 1k tokens; KnowLine 31B-Gen1 ties at 85.2, with Fasting GLIDE 1.2, H2O-Lightning-31B, and Perplexity Decider v1.1 27B rounding out the top five.
- Revised cost methodology: self-hosted open-weight models are now priced by GPU time (monthly-rented GPU, efficient engine) rather than chat list prices, since API providers also run cheap GPUs and optimized engines — list prices don't reflect real inference economics.
Jev-class systems are those within 2x the cost and median latency of the Jev v1.5 reference; capability score averages Intelligence and Calibration. 46 more open models were added to the rankings.
More from Infra
- Jensen Huang calls Nvidia's Megaspeed deals "normal" after exec's spying investigation report — pstAsiatech · 2026-10-30
- Intel guides weak on Q4, stock plunges 21.92% after hours — jd_pressman · 2026-10-30
- Why the 512GB M5 Ultra costs so much: 64 RAM channels need custom 8GB chips — two 256GB units may beat one — kingslyj · 2026-10-30
- Buying modded GPUs from Taobao: Reddit user maps the customs grey zones of freight forwarding — kh40tika · 2026-10-30
- 150,000 lines of PHP replaced by ~900 lines: dev ditches Coolify for self-built control plane — fekdaoui · 2026-10-30
- LFM3 runs fully offline on iPhone: 40 tok/s text, custom ExecuTorch export open-sourced — K4anan · 2026-10-30