FrontiersMind ships Lumma-Fev-v2: 4B and 27B with bidirectional attention, 0.1B and 0.6B to follow

Source · Model card: Lumma-fev

publicado (OK do Fabricio 30/10 BRT).

FrontiersMind published Lumma-Fev-v2 on 29 October 2026. Two models are out under Apache-2.0: Lumma-fev-27B-v2 (full fine-tune of Qwen3.8-27B) and Lumma-fev-4B-v2 (Qwen3.5-4B). A 0.1B on the company’s Nandi-Mini-150M base appeared in the collection on 30 Oct; the 0.6B is “coming soon”.

What changed. The 27B’s full-attention layers now read the prompt in both directions, while its Gated DeltaNet layers stay causal, so it needs lumma-fev 0.3 or later. Training data roughly tripled (~300K samples, from ~100K), and the models were trained with FrontiersMind’s own decisionformer. Interface is unchanged: noul and choice are trained, score is accepted but untested, context is 8,192 tokens, and lumma-fev-serve speaks /v1/systemone.

Vendor numbers (UnverifiedClaim). Average accuracy on eight public benchmarks: 87.3% for the 27B-v2 and 82.4% for the 4B-v2, against 81.2% for Jev 1.13.0 on the same rows. Jev still leads on BBH (91.1 vs 83.9) and JudgeBench (80.6 vs 77.4), and the blog says so. Pooled ECE 0.010 (27B) and 0.004 (4B) against 0.091 for Jev; an internal Indic set (400 rows, translated by Claude) shows 87.7% vs 79.7%. The launch post says the model is “Top 3 on the Decision Index, with the PR awaiting merge”.

What the boards show (30 Oct). The Decision Index 0.4.1 has only v1 (Lumma-fev-27B, Full 59.97, #50). No v2 row on the Decision Index, JevBench or Artificial Analysis yet, so the “Top 3” is a claim until the PR lands.

Score. The Lumma-fev page keeps its scores (5.1 / 6.1 / 22) until a board row exists.