1.7B TwIL-LM2 Specializes in Formal Logic, Beats 26B Giants
itsalreadyhardenough · reddit · 2026-08-15
TwIL-LM2 is a fine-tuned model (LoRA, 72M params) based on SmolLM2-1.7B, specialized solely for translating natural language into verifiable First-Order Logic, without general chat or reasoning capabilities.
Performance Highlight:
- On strict formal logic scoring (strict-7, no partial credit), TwIL-LM2 scored 0.2386, beating Qwen3-8B (0.2093) and Gemma-4-26B (0.2050).
- Larger models still win on "loose match" scoring, but the 1.7B model excels at emitting the exact formal representation required for downstream verification.
Use Case:
- Highlighted as an ideal "specialization" play: a small model doing one specific job well, slotted into a pipeline where a larger general model handles complex reasoning.
Note: Released under a Non-Commercial License.
Related event: 1.7B Logic-Tuned Model Outperforms Qwen3-8B on Formal Logic(2 posts)→
More from Models
- anon-1b: an open-weights Gemma fine-tune trained on 4chan greentext — rickasaurus · 2026-10-16
- Astra 6.1 reportedly delayed while OpenAI's frontier RL stays paused over safety concerns — haider1 · 2026-10-16
- Span-02 debuts: open-weight 'anytime reasoning' model that takes a deadline, not an effort level — HeyAmit_ · 2026-10-16
- Dev: Claude writes 'zero bugs' on healthcare apps, 10x token efficiency — repligate · 2026-10-16
- Shanghai AI Lab unveils Intern Lumina U2, a diffusion LLM unifying image, video and 3D tasks — Shanghai-AI-Laboratory · 2026-10-16
- This Week in Tech breaks down Gemini 4, OpenAI's Dots agents, and Anthropic's leaked IPO filing — LouMM · 2026-10-16