webAI CEO David Stout announced on X on November 4 (17:37 UTC) that the company is open-sourcing TwIL-8B. The official Hugging Face model card says it is post-trained from IBM Granite 4.2 8B with full-parameter SFT, GRPO/DAPO reinforcement learning and a length-aware RL stage, and ships in balanced and verbose variants. Self-reported scores include LiveCodeBench 81.6 (verbose), MATH-500 98.2 and IFBench 74.4, about 93% and 92% of the best frontier score, while GPQA Diamond is only 70.2. A 4.98 GiB Q4_K_M GGUF runs locally; the license is the webAI Non-Commercial License.
Key Takeaways
- ✓TwIL-8B has 8.79B parameters, is built on IBM's ibm-granite/granite-4.2-8b, has a 131,072-token context window and is English-only, text to text.
- ✓Self-reported scores: MATH-500 98.2, AIME 2025 88.3, LiveCodeBench 81.6 (verbose) and IFBench 74.4 (verbose); GPQA Diamond at 70.2 trails Qwen3.5-9B (81.7) clearly.
- ✓The balanced variant averages about 19.5k output tokens per AIME 2026 problem, about 38% fewer than Qwen3.5-9B's 31.2k, but still roughly 4x GPT-5.5.
- ✓All scores are webAI internal evaluations; competitor figures are quoted from public sources and were not re-run. SWE-bench, tool use and agentic tasks were not evaluated.
- ✓The license is the webAI Non-Commercial License 1.0, so it cannot be used commercially as is; the Q4_K_M GGUF is about 4.98 GiB and runs with llama.cpp.
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and the Industry Pain Point
Whether small models can approach frontier reasoning on local hardware is a long-running question for developers. webAI CEO David Stout posted on X that "modest intelligence is arriving faster than most people realize" and announced the open-sourcing of TwIL-8B, pitched as a model that runs on ordinary hardware such as laptops.
Architecture Highlights and Mechanism
According to the official model card, TwIL-8B is built on IBM Granite 4.2 8B (a 40-layer dense transformer). Full-parameter SFT is followed by GRPO/DAPO reinforcement learning to produce the verbose variant, then a length-aware RL stage produces the default balanced variant with shorter reasoning traces and lower cost per answer. Training pools were decontaminated with 13-gram matching against AIME, HMMT, GPQA, LiveCodeBench and others.
Benchmarks and Measured Results
Self-reported scores: MATH-500 98.2, LiveCodeBench 81.6 and IFBench 74.4 (the last two on the verbose variant), about 93% and 92% of the best score among GPT-5.5, Claude Opus 4.8 and Grok 4.3. Knowledge tasks lag clearly, with GPQA Diamond at 70.2 and HLE at 12.6. These are webAI internal evaluations; competitor figures were quoted from public sources without re-running, and there is no third-party reproduction yet.
Getting Started for Developers
Load webAI-Official/TwIL-8B with transformers (balanced by default, subfolder="verbose" for the verbose variant), or run the 4.98 GiB Q4_K_M GGUF with llama.cpp. webAI recommends a large generation budget of tens of thousands of tokens, since truncated reasoning scores far worse. The license is non-commercial, and there is no extra safety or tool-use training.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.