Reported News
webAI TwIL-8B: an 8.8B reasoning model that reports 98.2 on MATH-500, non-commercial weights
webAI released TwIL-8B on November 4, 2026, an 8.79B-parameter reasoning model post-trained from IBM's Granite 4.2 8B that its model card says scores 98.2 on MATH-500. The weights are on Hugging Face under the webAI Non-Commercial License, with a 4.98 GiB Q4_K_M GGUF for local use.
- Changes reported
- 6
- listed below, each from a source we read
- We tested
- Not tested yet
- reported from sources; we did not run it ourselves
- Gaps we flag
- 4
- things we could not verify
- Sources
- 1
- 1 primary, linked below
What changed
| # | What the source says | Status |
|---|---|---|
| 01 | New open-weight model: an 8.79B-parameter dense decoder-only model post-trained from ibm-granite/granite-4.2-8b, with a 131,072-token context window inherited from the base model. Text in, text out, English only. | Reported |
| 02 | Two variants share one architecture and tokenizer: balanced (the default, shorter reasoning traces) and verbose (longer reasoning for hard contest problems). Both ship as full merged models, not LoRA adapters. | Reported |
| 03 | webAI reports 98.2 on MATH-500, 88.3 on AIME 2025 and 81.6 on LiveCodeBench (verbose variant), the top scores in its comparison against Qwen3.5-9B, Gemma 4 12B and Gemma 4 31B. | Reported |
| 04 | Knowledge is the weak spot by the card's own numbers: GPQA Diamond is 70.2, 11.5 points behind Qwen3.5-9B and 14.1 behind Gemma 4 31B. | Reported |
| 05 | The card reports about 19.5k output tokens per AIME 2026 problem, 38% fewer than Qwen3.5-9B's 31.2k, at 4.2 points lower accuracy. | Reported |
| 06 | Weights are 16.38 GiB per variant in bf16, with GGUF builds for llama.cpp down to a 4.98 GiB Q4_K_M. The license is the webAI Non-Commercial License ver. 1.0. | Reported |
reported by the sourcemeasured or run by usnot verified
Primary source: Hugging Face model card: webAI-Official/TwIL-8B. We are the fast, independent index; the vendor’s own page is the source of record for this release.
In the source’s words
“It is level with Qwen3.5-9B on contest math while using fewer tokens per problem.”
“A short budget truncates the reasoning and scores far worse.”
What we have not verified
- We did not run TwIL-8B; every TwIL-8B score is webAI's own internal high-effort pass@1 run, and competitor figures were copied from other published sources, not re-run.
- The card says the LiveCodeBench version and date window for TwIL-8B's run were not recorded, while competitor scores are v6, so that gap is not like for like.
- Rows marked as verbose-only (LiveCodeBench, MMLU-Pro, HLE, IFBench and others) have not been run on the default balanced variant.
- The model card has no publish timestamp; the November 4, 2026 date comes from the card's last update and press coverage the same day.
This brief is Reported: written from the sources below, not from running the release ourselves.
Sources
- Hugging Face model card: webAI-Official/TwIL-8Bprimary source · fetched 2026-11-04
Related
Spotted an error? See our corrections policy or email the address on the contact page. Markdown version for agents: /news/twil-8b-open-weight-reasoning-model.md
Want these briefs in your Google Top Stories? Add Agentic Weekly as a preferred source on Google.
Newsletter
The Weekly Verdict. Free, every Monday.
A free email every Monday: the week's few big AI releases, then 5 short use cases for the new models, each with a prompt you can paste.
Free. Double opt-in. One click to unsubscribe.