webAI-Official/TwIL-8B · Hugging Face
This is the Hugging Face model card for TwIL-8B, an 8.8B-parameter model fine-tuned from IBM's Granite 4.2 for math, code and instruction following. The card covers how to run it through Transformers, llama.cpp, vLLM, Ollama and several local apps, plus benchmark results against similar-sized open models.
- Scores 98.2 on MATH-500 and 81.6 on LiveCodeBench (verbose variant), the top results in its comparison group, with a Q4_K_M quantized GGUF about 5 GB so it runs on modest hardware.
- Trails on knowledge-heavy benchmarks: GPQA Diamond is 11.5 points behind Qwen3.5-9B, and the model is not a general knowledge model.
- Uses about 38% fewer output tokens per AIME 2026 problem than Qwen3.5-9B, though it is still far more verbose than frontier models.
TwIL-8B is an 8.8B-parameter reasoning model built on IBM's granite-4.2-8b, fine-tuned with full-parameter supervised training and then pushed further with GRPO and DAPO reinforcement learning. It ships in two variants that share architecture and tokenizer: a balanced default and a verbose one. The headline numbers are for the balanced variant, though a few benchmarks (marked with a dagger) were only run on verbose so far. The model is aimed at math, code, and instruction following rather than broad knowledge.
The benchmark claims are strongest where you'd hope a reasoning model would be. It posts 98.2 on MATH-500, ahead of Qwen3.5-9B at 96.6 and Gemma 4 31B at 87.9, and 88.3 on AIME 2025, edging out Gemma 4 31B. On LiveCodeBench it scores 81.6 (verbose), about 16 points above Qwen3.5-9B and roughly 1.6 above Gemma 4 31B. Instruction following is also solid: IFBench at 74.4 is nearly 10 points over Qwen3.5-9B, though it sits a bit behind Gemma 4 31B. Contest math results are mixed. Across four HMMT and AIME contests, it averages about even with Qwen3.5-9B, winning some and losing others, including a 4.2-point deficit on AIME 2026.
The weak spots are knowledge-heavy tasks. GPQA Diamond, MMLU-Pro and HLE all trail the comparison group, which is consistent with a small model optimized for reasoning over recall. Keep in mind that competitor scores are republished figures rather than re-runs, so small gaps deserve some skepticism. For running it, the page gives ready recipes for Transformers, llama.cpp (including a Q4_K_M GGUF), vLLM, SGLang, Ollama, LM Studio and several agent frameworks, so getting it running locally is straightforward.