webAI Releases TwIL-8B, an 8.8B Reasoning Model Scoring 98.2 on MATH-500

webAI released TwIL-8B, an 8.79B dense reasoning model post-trained from IBM's Granite 4.2 8B. Weights are on Hugging Face under webAI's Non-Commercial 1.0 license. The pitch is cost per task: a 4.98 GiB Q4_K_M GGUF for CPU or modest-VRAM GPUs, 131K context inherited from the base, and two variants — balanced (shorter traces) and verbose (harder contest problems).

Self-reported high-effort pass@1 scores put TwIL-8B at 98.2 on MATH-500 and 81.6 on LiveCodeBench (verbose). Against Qwen3.5-9B it goes 7–6 across 13 shared benchmarks; against Gemma 4 31B, roughly 3.5× its size, it goes 3–11. Knowledge is the soft spot — GPQA Diamond at 70.2 trails Qwen3.5-9B by 11.5 points.

We're interested in small reasoning models that don't pretend to be frontier generalists. TwIL-8B's training story is clean: full-parameter SFT on reasoning traces, then GRPO/DAPO RL for the verbose variant, then length-aware RL that rewards shorter traces against per-domain references for the balanced variant. Decontamination used 13-gram matching against AIME, HMMT, MATH-500, GPQA and other evals — the right hygiene to claim, even if we always want third-party re-runs.

The token-efficiency angle is the part I'd put on air. On AIME 2026, TwIL-8B averages about 19.5k output tokens versus 31.2k for Qwen3.5-9B — 38% fewer — while scoring 4.2 points lower. That's a real cost trade for local or budget inference. Frontier models still crush that budget (webAI cites ~5k tokens for GPT-5.5 on the same set), so the gap to closed models remains, but the open mid-size race is about tokens spent per correct answer, not just peak score.

License is the friction. Apache 2.0 peers like Qwen3.5-9B and Gemma 4 can ship in products; TwIL-8B cannot without a commercial deal. For research labs, indie builders, and classroom use, that may be fine. For anyone building a paid app, it's a stop sign until the paperwork changes.

I'm watching whether "balanced vs verbose" becomes a standard shipping pattern. Giving operators a short-trace default and a long-trace contest mode is more honest than one checkpoint that overthinks every chat. The model card's example max_new_tokens of 32,768 is also a quiet warning: starve the generation budget and the reasoning scores collapse.

Useful AI when the evidence supports it — a locally runnable math/code specialist at this size is a win. Eyes open on the knowledge gap and the non-commercial leash.

TwIL extends webAI's earlier 1.7B and 3B formal-logic models from August 2026. The architecture is a dense decoder-only Transformer (40 layers, hidden 4096, 32 attention heads / 8 KV heads) that emits a thinking block before each answer. No hosted Hugging Face inference provider is listed yet.