Adapter-LoRA on Evo2 + Qwen3 beats Carbon-X DNA model with 24h of cheap compute
danielmckinn0n · x · 2026-10-27
Inspired by Hugging Face's Carbon-X DNA model, the author reused the early vision-language adapter recipe (LLaVA-style) and let Codex take over his 5090 GPU:
- Frozen Evo2 7B (Arc Institute) as DNA encoder, Qwen3-14B as the reasoning backbone, plus a task-specific LoRA
- With only 24h of cheap compute and occasional human intervention, the adapter-LoRA approach reportedly beats Carbon-X across the board and matches Astra on his self-built tasks
- Screenshot evals show 55.6% accuracy vs Carbon-X's 28.9% and Astra subagents' 47.3%
- He is running BioReason, GenomeQA and other Carbon-X benchmarks overnight and plans a blog post, pushing the minimal-compute adapter route further
More from Research
- Looped transformer analysis pins down single neurons for non-duality, phonetics, and other concepts — burny_tech · 2026-10-27
- LLMs negotiate well in deterministic settings but only 3 of 9 pairings close deals under high stochasticity — davidad · 2026-10-27
- Single-neuron probes of Qwen3.5 surface a cast of hidden personas, from Naruto to cartoon mascots — burny_tech · 2026-10-27
- Sakana AI RSI Lab showcases Schmidhuber keynote, Japan-built foundation model Sawara, and open-model momentum — Stefania_druga · 2026-10-27
- ETH proposers lead open review of solver claim; exponent shrink factor ~0.67 at a billion variables — AIFlow_ML · 2026-10-27
- PSI's AI research system cracks five open quantum information problems, with Lean proofs — IgorCarron · 2026-10-27