Gemma-4-31B fine-tune tops Decision Index, validating non-causal attention
multimodalart · x · 2026-10-12
Multimodal community author multimodalart notes that a fine-tune of Gemma-4-31B-it (Mapika's Decider 31B) ties for #1 on Decision Index 0.3.1, while the stock model ties #3 using the AnyJev inference technique — suggesting the base model is a strong ground for Jev-like models. The #1 and #2 spots both use Perplexity's Decider v1.1 non-causal attention architecture across independently trained tunes, validating the design; the author speculates Gemma-4-31B with non-causal attention could be a monster. Simpler tricks like reading the same input twice (as with Shisa) also land near the top.
Related event: Decision Index 0.3.1: vLLM d3 and Mapika Decider 31B Tie for First(2 posts)→
More from Models
- PlaySuite benchmarks 14 open models across 5,734 video games — cgmsnoek · 2026-10-12
- Cheap model finished the job in half the time at 80% accuracy vs frontier — victor_explore · 2026-10-12
- Use small models for grunt work, save SOTA for orchestration, says agentic engineering practitioner — petergyang · 2026-10-12
- DeepSeek model claims to be Claude then ChatGPT in back-to-back replies — Div_pradeep · 2026-10-12
- Anthropic rumored to be testing a new model, suspected early Opus 5.5 checkpoint — iamfakhrealam · 2026-10-12
- Open-source leaderboard whatmodel.sh shows the best local models your machine can run — jakecoolguy · 2026-10-12