webAI-Official/TwIL-8B
AI model · https://huggingface.co/webAI-Official/TwIL-8B
First seen by Oriel on .
AI summary
TwIL-8B is an 8.8B model post-trained from Granite 4.2 for contest math, code, and instruction following, with balanced and verbose variants trading shorter reasoning for longer traces.
- The model combines full-parameter supervised fine-tuning with GRPO/DAPO reinforcement learning; length-aware RL produces the balanced variant.
- It reports 98.2 on MATH-500 and 81.6 on LiveCodeBench, while trailing larger models on knowledge benchmarks such as GPQA Diamond.
- Balanced uses about 19.5k output tokens per AIME 2026 problem, 38% fewer than Qwen3.5-9B, but with 4.2 points lower accuracy.
- The 131,072-token-context model has 8.79B parameters and ships in bf16 and GGUF formats; its Q4_K_M build is 4.98 GiB.
- Several results are from the verbose variant, competitor scores were not rerun, and safety, preference, tool-use, and agentic tasks were not evaluated.
#large language models #reinforcement learning #benchmarking #gguf
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Likes
Total likes reported by the source.
- Hugging Face Models: 420, latest observation ; +322 over observed period since
Downloads (30d)
Rolling 30-day downloads. A change in this value is not new downloads during the observed period.
- Hugging Face Models: 61,393, latest observation ; +57,830 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Models | webAI-Official/TwIL-8B | https://huggingface.co/webAI-Official/TwIL-8B Read discussion |