Team mid-trains Qwen3.5 14B to frontier level on DNA — and releases the recipe
lvwerra · x · 2026-10-22
lvwerra's team claims they mid- and post-trained Qwen3.5 14B at a fraction of the cost, without distillation, to beat every other open model and match the best frontier models.
Key points:
- The training target is understanding and reasoning over DNA sequences — not a toy task. Since the model learns both natural language and DNA, you can "chat to your DNA": ask about mutation impacts or generate small sequences.
- They argue bio is the next AI frontier and expect frontier labs to shift from math and cyber toward bio; researchers should have access to capable models.
- The full training recipe is being released for reproduction.
Related event: Hugging Face Open-Sources Carbon-X, a 14B DNA-Language Model(19 posts)→
More from Models
- Similarweb: Grok Drew 40.91% New-User Traffic Share in September, Highest Among Major AI Tools — XFreeze · 2026-10-22
- Rumor: 'Gemini 4 Carbon' updated checkpoint spotted in cryptic Carbon vs Argon teaser — lyraxana · 2026-10-22
- Recursive Bets on RSI: Its Eureka Machine Built the Flint Coding Model in Months — mikeflache · 2026-10-22
- 0.9B Model Transcribes a 4-Person Meeting on a MacBook With 0.08% Speaker Confusion — Binary_orchid · 2026-10-22
- Fireworks trains Meta's Muse Glimmer on Kalshi outcomes: +11.2pp accuracy over Jev — sophiamyang · 2026-10-22
- Gemini CLI PR reveals unreleased gemini-4-argon model with 1M context window — mr8lu · 2026-10-22