webAI's TwIL-8B Beats Frontier Models on Math and Code at 5 GiB
webAI released TwIL-8B , an 8.79B reasoning model post-trained from IBM's Granite 4.2 8B. Scores 98.2 on MATH-500 and 81.6 on LiveCodeBench, topping its weight class. Reaches 88 to 93% of frontier model scores on contest math, code and instruction following.
Takeaways
- webAI released TwIL-8B , an 8.79B reasoning model post-trained from IBM's Granite 4.2 8B.
- Scores 98.2 on MATH-500 and 81.6 on LiveCodeBench, topping its weight class.
- Reaches 88 to 93% of frontier model scores on contest math, code and instruction following.
- Uses 38% fewer output tokens than Qwen3.5-9B on AIME 2026 thanks to length-aware RL.
- Ships balanced and verbose variants; Q4_K_M GGUF is just 4.98 GiB for local inference.
- Weights are open but licensed non-commercial only; trails frontier on GPQA, MMLU-Pro and HLE knowledge benchmarks.
webAI has released TwIL-8B , an 8.79-billion-parameter reasoning model post-trained from IBM’s Granite 4.2 8B. The company reports scores of 98.2 on MATH-500 and 81.6 on LiveCodeBench for the verbose variant, placing it near GPT-5.5 and Claude Opus 4.8 on those specific tests. The downloadable weights carry a noncommercial license.
The Q4_K_M artifact occupies 4.98 GiB, making local deployment practical on consumer hardware with sufficient memory. Its benchmark profile favors math, code, and instruction following, while knowledge-heavy evaluations remain a clear weakness. Runtime memory exceeds the weight-file size because inference also requires a key-value cache, activations, and framework overhead.
| Detail | TwIL-8B |
|---|---|
| Parameters | 8.79 billion |
| Base model | IBM Granite 4.2 8B |
| Variants | Balanced and verbose |
| Maximum context | 131,072 tokens |
| BF16 checkpoint | 16.38 GiB |
| Q4_K_M artifact | 4.98 GiB |
| License | Noncommercial |
Math and code carry the scorecard
webAI’s published comparison places TwIL-8B ahead of downloadable-weight models ranging from a similar size to roughly 3.5 times larger on MATH-500 and LiveCodeBench. It also leads Qwen3.5-9B by 9.9 points on IFBench, which measures compliance with detailed instructions.
| Benchmark | Reported result | What it measures |
|---|---|---|
| MATH-500 | 98.2 for the verbose variant | Competition-style mathematical problem solving |
| LiveCodeBench | 81.6 for the verbose variant | Code generation against recent problems and tests |
| IFBench | 9.9 points above Qwen3.5-9B | Precise instruction following |
| HLE | 26% of the best frontier score in webAI’s comparison | Broad, difficult academic knowledge and reasoning |
TwIL-8B reaches 93% of the best frontier result on LiveCodeBench, 92% on IFBench, and between 88% and 90% on three of four contest-math sets in webAI’s comparison. Its results fall further behind on GPQA Diamond, MMLU-Pro, and HLE, which require broader factual and academic knowledge. The pattern supports a narrow conclusion: targeted post-training can concentrate a small model’s capacity on verifiable tasks.
Cross-vendor benchmark scores can shift with prompt templates, sampling settings, generation limits, grading methods, and benchmark versions. These figures come from webAI’s model card and need independent replication before they can support purchasing or architecture decisions. They establish task-specific competitiveness rather than general parity with frontier systems.
Output length shapes serving cost
On AIME 2026, the balanced variant uses 38% fewer output tokens than Qwen3.5-9B. On HMMT February 2026, it uses 19% fewer tokens while scoring three points lower. Shorter generations reduce inference time and GPU occupancy when a deployment is limited by decoding throughput.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
