Playbooks · Coding tasks
TwIL-8B: a local model for math and coding
Written November 9, 2026
TwIL-8B is a local 8.8B reasoning model for math and coding. What its card claims, where it trails larger models, and the licence check first.
The problem it solves
A developer who wants a local reasoning model for math and coding has a new 8B candidate. TwIL-8B, from webAI, is post-trained from IBM's Granite 4.2 8B for contest math, code, and instruction following, and it ships as open weights with a 4.98 GiB quantized build for CPU or modest GPUs. The local-model cell is crowded — Qwen3.5-9B and Gemma 4 12B and 31B sit on the same shelf — so the question is not whether TwIL-8B runs locally, but whether its numbers earn it a place there.
What it claims
TwIL-8B has 8.79 billion parameters, a 131,072-token context window, and two variants sharing one architecture and tokenizer: balanced (the default, shorter reasoning traces) and verbose (longer traces for hard contest-style problems). The model writes a <think> block before each answer. Full bf16 weights are 16.38 GiB; the Q4_K_M GGUF build is 4.98 GiB and runs under llama.cpp with a command the card gives:
llama-cli -hf webAI-Official/TwIL-8B --hf-file TwIL-8B-Q4_K_M-balanced.gguf -cnv -c 40960 -n 32768
The card's headline figures, all webAI's own and none independently rerun: MATH-500 98.2, against 96.6 for Qwen3.5-9B and 87.9 for Gemma 4 31B; AIME 2025 88.3; LiveCodeBench 81.6, against 65.6 for Qwen3.5-9B and 80.0 for Gemma 4 31B; IFBench 74.4, 9.9 points above Qwen3.5-9B. Against the frontier it reaches 93% of the best score on LiveCodeBench, 92% on IFBench, and 90% on HMMT Nov 2025. It also uses fewer tokens than its closest open peer: about 19.5k output tokens per AIME 2026 problem against 31.2k for Qwen3.5-9B, 38% fewer, for 4.2 fewer points.
Where it trails
The card is candid about the weak side, and it is the side that decides most purchases. TwIL-8B is not a knowledge model: GPQA Diamond 70.2 trails Qwen3.5-9B by 11.5 points and Gemma 4 31B by 14.1, and MMLU-Pro and HLE follow the same pattern. On the hardest contest set, HMMT Feb 2026, it scores 68.2 — 69% of the best frontier score, the outlier in an otherwise 88–93% picture.
Three caveats in the card deserve attention before any of the headline numbers are quoted. The LiveCodeBench, IFBench, and several other rows were measured on the verbose variant and have not yet been run on balanced, the default. The card's own LiveCodeBench version and date window are not recorded, so the 16-point gap over Qwen3.5-9B should be confirmed before it is quoted. And SWE-bench, SuperGPQA, tool use, and agentic tasks were not evaluated at all — there is no evidence here about how the model does on the agent work a coding assistant actually performs.
The check before you run it
First, the licence: TwIL-8B is released under the webAI Non-Commercial License v1.0. The weights are open, but use is for research and study, not paid products. That rules it out for anything commercial before any benchmark matters.
Second, the budget: give it a large generation budget. Balanced averages about 19.5k tokens per AIME 2026 problem and 33k per HMMT Feb 2026 problem, and verbose writes more. A short budget truncates the reasoning and scores far worse — the card is explicit about that. There is no hosted inference provider serving the model, and all reported scores used the bf16 weights, so expect small deviations at Q4_K_M.
Verdict
Worth a local run for math and code reasoning where a non-commercial licence fits — the card figures lead its size class on MATH-500 and LiveCodeBench, and the 4.98 GiB build makes the test cheap. Not a general assistant, and not for knowledge-heavy work: the GPQA gap is wide, tool use and agentic tasks were never evaluated, and no independent run has confirmed any of the card's numbers yet.
Limits
- Every benchmark figure on this page is webAI's own. Competitor scores are published figures that were not rerun, and no independent evaluation of TwIL-8B exists.
- Rows marked as verbose-variant measurements have not yet been run on balanced, the default variant — including LiveCodeBench 81.6, the headline code figure.
- The card records no LiveCodeBench version or date window for its own run, and the AIME and HMMT sets have only 30–33 problems, so a single problem is worth about 3 points.
- Tool use and agentic tasks were not evaluated, and there is no safety, preference, or tool-use training beyond what Granite 4.2 carries.
- Use is limited by the webAI Non-Commercial License v1.0: research and study, not paid products.
Related playbooks
- MAI-Code-1.1-Flash: local coding on RTX Spark
- Ripple-65B: the $32M open model topping coding tests
- DeepSeek's cheap coding model changed under the same name: retest Flash 1017 before you rely on it
- Coding-agent traces in advisory fixes
Sources
- webAI-Official/TwIL-8B — Hugging Face
- webAI Releases TwIL-8B, an 8.8B Reasoning Model Scoring 98.2 on MATH-500 — MarkTechPost
Back to: Coding playbooks hub · Also: MAI-Code-1.1-Flash: local coding on RTX Spark · Foundry Local runs Qwen3.8-27B on your GPU · Run LongCat-2.5 in Claude Code without the 200K cap