FrontiersMind open-sources Lumma-fev-27B, a decision model scoring options in one forward pass
kalyan_kpl · x · 2026-10-14
FrontiersMind has open-sourced Lumma-fev-27B, its largest "decision model": instead of generating a label to parse, it scores the options you provide in a single forward pass. Built on Qwen3.8-27B with full fine-tuning (no LoRA, no frozen layers) on 100K decision samples, it supports English plus three Indic languages: Hindi, Tamil, and Kannada.
On a held-out set of 4,454 decisions, Lumma-fev-27B is correct 90.7% of the time vs 83.9% for TypeSafe Jev, with better-calibrated confidence; Jev remains ahead on reasoning benchmarks like BBH and WinoGrande. The interface is unchanged, supporting noul (yes/no gate), choice (up to 255 options), and score (ordered levels). Weights are available on Hugging Face under Apache 2.0.
More from Models
- Providers quietly used cheap classifier routing for years while charging full LLM prices — sull · 2026-10-14
- TokenGauge price index: same LLM varies up to 85% across providers and gateways — const_reborn · 2026-10-14
- Gemini 4 Argon writes 27x PyTorch-speedup GPU megakernel, ranks 3rd on KernelBench-Mega — scaling01 · 2026-10-14
- Claude Haiku 5.5 Lands on Runware API at $0.10/M Input Tokens — aziz4ai · 2026-10-14
- Gemini 4 Argon is good, but access is confusing: unavailable in Antigravity — dreamwieber · 2026-10-14
- NVIDIA's Nemotron 3.5 Bolt 74B MoE fits 24GB via 2-bit quantization — CulturalKing5623 · 2026-10-14