webAI发布8.8B推理模型TwIL-8B,MATH-500得98.2分
webAI 发布了 TwIL-8B,这是一个 8.8B 的推理模型,面向数学、代码和指令遵循。TwIL-8B 推理模型基于 IBM 的 Granite 4.2 8B 进行后训练。权重以非商业许可在 Hugging Face 上发布。webAI 的发布文章将卖点围绕每任务成本展开,而非原始规模。4-bit 版本为 4.98 GiB,面向 CPU 或显存有限的 GPU 推理。
TwIL-8B 是 webAI 推出的稠密仅解码器推理模型。它有 40 层,隐藏维度 4096,32 个注意力头、8 个 KV 头。输入和输出仅支持英文文本。模型在每次回答前会写一个 <think> 块。
它有两个变体,架构和分词器相同。Balanced 是默认版本,产生的推理轨迹更短。Verbose 推理更长,面向高难度竞赛类题目。
TwIL-8B 扩展了 webAI 的 TwIL 系列。首批成员是 2026 年 8 月发布的 1.7B 和 3B 形式逻辑模型。
webAI 采用了 3 个后训练阶段:
训练数据池通过 13-gram 匹配针对 AIME、HMMT、MATH-500、GPQA 及其他评测集进行了去污染处理。
所有 TwIL-8B 分数均为自报,high-effort pass@1、启用思考。标记 † 的分数仅在 verbose 变体上测得。
在 13 项共同基准测试中,TwIL-8B 对阵 Qwen3.5-9B 取得 7-6 的成绩。对阵规模约为其 3.5 倍的 Gemma 4 31B,成绩为 3-11。
是的,与它最接近的开放同类模型相比。在 AIME 2026 上,TwIL-8B 平均每道题输出约 19.5k Token。Qwen3.5-9B 使用 31.2k,因此 TwIL-8B 少用 38%。但它的得分低 4.2 分。在 2026 年 2 月的 HMMT 上,差距为 32.9k 对 40.6k,即少用 19%。
前沿模型仍然精简得多。webAI 引用 GPT-5.5 在 AIME 2026 上每道题约 5k Token 的数据。webAI 表示,各家的评测框架与分词器并不相同,因此这只能作为方向性参考。
TwIL-8B 的数据来自模型卡。† = 冗长版本。除非另有标注,所有得分均由各厂商自行报告。
from_pretrained("webAI-Official/TwIL-8B") 加载的是平衡版本。传入 subfolder="verbose" 则加载冗长版本。模型卡列出了适用于 llama.cpp 的 GGUF 构建版本:
要给它足够大的生成预算。模型卡中的示例设置了 max_new_tokens=32768。预算过短会截断推理过程,得分会差得多。所有已报告的得分都使用了 bf16 权重,因此在 Q4_K_M 下预计会有小幅下降。目前 Hugging Face 上还没有托管推理服务商提供该模型。
想与我们合作推广你的 GitHub 仓库、Hugging Face 页面、产品发布或网络研讨会等?欢迎联系我们
文章《webAI 发布 TwIL-8B:一款在 MATH-500 上得分 98.2 的 8.8B 推理模型》首发于 MarkTechPost。
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(webAI Releases TwIL-8B, an 8.8B Reasoning Model Scoring 98.2 on MATH-500)
webAI has released TwIL-8B, an 8.8B reasoning model for math, code and instruction following. The TwIL-8B reasoning model is post-trained from IBM’s Granite 4.2 8B. Weights are on Hugging Face under a non-commercial license. webAI’s launch post frames the pitch around cost per task, not raw scale. The 4-bit build is 4.98 GiB, sized for CPU or modest-VRAM GPU inference.
TL;DR
- Size: 8.79B parameters, dense. 131,072-token context, inherited from the base model.
- Runs on: CPU or modest-VRAM GPU via a 4.98 GiB Q4_K_M GGUF. Full bf16 weights are 16.38 GiB.
- Performance: Leads its size class on math and code, trails on knowledge benchmarks.
- Best: MATH-500 at 98.2, the top score in webAI’s comparison.
- Worst: GPQA Diamond at 70.2, 11.5 points behind Qwen3.5-9B. HLE is 12.6 (verbose variant).
What is TwIL-8B?
TwIL-8B is a dense decoder-only reasoning model from webAI. It has 40 layers, hidden size 4096, and 32 attention heads with 8 KV heads. Input and output are English text only. The model writes a <think> block before each answer.
It ships in 2 variants with the same architecture and tokenizer. Balanced is the default and produces shorter reasoning traces. Verbose reasons longer and targets hard contest-style problems.
TwIL-8B extends webAI’s TwIL family. The first members were 1.7B and 3B formal-logic models released in August 2026.
How was TwIL-8B trained?
webAI used 3 post-training stages:
- Full-parameter SFT on reasoning traces for math, STEM and instruction following.
- GRPO/DAPO reinforcement learning on math, code, STEM and instruction prompts. This produces the verbose variant.
- Length-aware RL with a reward that favors shorter traces against per-domain length references. This produces the balanced variant.
Training pools were decontaminated with 13-gram matching against AIME, HMMT, MATH-500, GPQA and other evals.
How does TwIL-8B perform on benchmarks?
All TwIL-8B scores are self-reported, high-effort pass@1 with thinking enabled. Scores marked † were measured on the verbose variant only.
- Math: TwIL-8B scores 98.2 on MATH-500. Qwen3.5-9B lists 96.6 on a third-party leaderboard. Gemma 4 31B scored 87.9 in Red Hat’s lighteval run. On AIME 2025, TwIL-8B posts 88.3 against 87.1 for Gemma 4 31B. On AIME 2026 it scores 88.3, behind Qwen3.5-9B at 92.5 and Gemma 4 31B at 89.2.
- Code: LiveCodeBench† is 81.6, against 65.6 for Qwen3.5-9B and 80.0 for Gemma 4 31B. webAI notes its LiveCodeBench version was not recorded. Competitor scores are v6.
- Instruction following: IFBench† is 74.4, 9.9 points above Qwen3.5-9B at 64.5. IFEval is level, at 91.8 vs 91.5.
- Knowledge: This is the weak spot. GPQA Diamond is 70.2, against 81.7 for Qwen3.5-9B and 84.3 for Gemma 4 31B. MMLU-Pro† is 76.5 vs 82.5 for Qwen3.5-9B.
Across 13 shared benchmarks, TwIL-8B goes 7-6 against Qwen3.5-9B. Against Gemma 4 31B, about 3.5x its size, it goes 3-11.
Does TwIL-8B use fewer reasoning tokens?
Yes, against its closest open peer. On AIME 2026, TwIL-8B averages about 19.5k output tokens per problem. Qwen3.5-9B uses 31.2k, so TwIL-8B uses 38% fewer. It scores 4.2 points lower. On HMMT Feb 2026 the gap is 32.9k vs 40.6k, or 19% fewer.
Frontier models remain far leaner. webAI cites about 5k tokens per AIME 2026 problem for GPT-5.5. webAI states that harnesses and tokenizers differ, so treat this as directional.
TwIL-8B vs Qwen3.5-9B vs Gemma 4
| Metrics | TwIL-8B | Qwen3.5-9B | Gemma 4 12B | Gemma 4 31B |
|---|---|---|---|---|
| Parameters | 8.79B dense | 9B (language model) | 11.95B dense | 30.7B dense |
| Context | 131,072 | 262,144 native | 256K | 256K |
| Modality (input) | Text | Text, image, video | Text, image, audio | Text, image |
| License | webAI Non-Commercial 1.0 | Apache 2.0 | Apache 2.0 | Apache 2.0 |
| Smallest official local build | 4.98 GiB (Q4_K_M) | Not disclosed | Not disclosed | Not disclosed |
| MATH-500 | 98.2 | 96.6 (3rd party) | Not disclosed | 87.9 (Red Hat) |
| AIME 2026 | 88.3 | 92.5 (MathArena) | 77.5 | 89.2 |
| LiveCodeBench | 81.6 † | 65.6 (v6) | 72.0 (v6) | 80.0 (v6) |
| GPQA Diamond | 70.2 | 81.7 | 78.8 | 84.3 |
| MMLU-Pro | 76.5 † | 82.5 | 77.2 | 85.2 |
TwIL-8B figures from the model card. † = verbose variant. All scores self-reported by each vendor unless marked.
How do you run TwIL-8B locally?
from_pretrained("webAI-Official/TwIL-8B") loads the balanced variant. Pass subfolder="verbose" for the verbose one. The model card lists GGUF builds for llama.cpp:
- Q4_K_M, 4.98 GiB, for both variants
- Q8_0, 8.70 GiB, verbose only
- F16, 16.38 GiB, verbose only
Give it a large generation budget. The card’s example sets max_new_tokens=32768. A short budget truncates reasoning and scores far worse. All reported scores used bf16 weights, so expect small drops at Q4_K_M. No hosted inference provider serves the model on Hugging Face yet.
Key Takeaways
- TwIL-8B is an 8.79B dense reasoning model post-trained from IBM Granite 4.2 8B.
- It posts 98.2 on MATH-500 and 81.6† on LiveCodeBench.
- It uses 38% fewer tokens than Qwen3.5-9B on AIME 2026.
- Knowledge is weak: GPQA Diamond trails Qwen3.5-9B by 11.5 points.
- Weights are open but licensed for non-commercial use only.
Check out the model card and weights on Hugging Face and webAI’s announcement on X. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us
The post webAI Releases TwIL-8B, an 8.8B Reasoning Model Scoring 98.2 on MATH-500 appeared first on MarkTechPost.
正文保留英文原文,标题与摘要由机器翻译。
原文发布于 MarkTechPost(webAI Releases TwIL-8B, an 8.8B Reasoning Model Scoring 98.2 on MATH-500),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。