webAI open-sources TwIL-8B, an 8.8B reasoning model it says reaches 93% of the frontier on LiveCodeBench
Post-trained from IBM's Granite 4.2 8B, it posts 98.2 on MATH-500 and 81.6 on LiveCodeBench in webAI's own tests but trails on knowledge benchmarks, and ships under a non-commercial licence.
Why it matters
TwIL-8B is the latest sign that small open models are closing in on frontier systems on narrow reasoning tasks: webAI says it beats Qwen3.5-9B by 16 points on LiveCodeBench and uses 38% fewer tokens per AIME 2026 problem. The gains are concentrated in maths and code; it trails Qwen3.5-9B by 11.5 points on GPQA Diamond and has no safety tuning beyond what Granite 4.2 carries. It fits in 4.98 GiB as a quantised file, so it runs on ordinary laptops. Watch for independent replication of the scores.
Key facts
- 8.79B parameters, post-trained from IBM Granite 4.2 8B
- MATH-500 98.2, AIME 2025 88.3
- LiveCodeBench 81.6, 93% of the best frontier score
- GPQA Diamond 70.2, 11.5 behind Qwen3.5-9B
- 4.98 GiB Q4_K_M GGUF
- webAI non-commercial licence