What is TwIL-LM3-Pro?
TwIL-LM3-Pro is a 3.66-billion-parameter model from webAI, the Austin company that sells on-device AI, announced on 30 September 2026. It is IBM's Granite 4.2 3B taken through LoRA fine-tuning on synthetic formal-logic data, multipath distillation, checkpoint merging and an entropy-weighted GRPO reinforcement-learning stage. It reads and writes English text and reasons in a think block before answering.
It is built for formal logic: deciding whether a conclusion follows from premises, inducing rules from examples, translating statements into first-order logic or semantic parses, and critiquing Lean statements. The weights are on Hugging Face in BF16 and GGUF from 1.98 GiB (Q4_0) to 6.82 GiB (F16), under the webAI Non-Commercial License 1.0, which allows research and education only.
How much does TwIL-LM3-Pro cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | Non-commercial research and education only |
| Commercial use | Not published | Not allowed under the published licence |
Free weights for non-commercial research and education under the webAI Non-Commercial License 1.0. No hosted API or commercial licence price published.
Is TwIL-LM3-Pro worth it?
TwIL-LM3-Pro is a narrow tool, and inside its niche the numbers webAI publishes are good. On webAI's own formal-logic composite it scores 0.5539, up from 0.4313 for the Granite base and roughly level with Qwen3-8B's 0.5336, a gap the model card itself calls within noise. It beats VibeThinker-3B (0.4118) and Qwen3.5-4B (0.4466) on the same composite, and its six-task logic average of 0.5389 edges past gpt-oss-120b's 0.5192, though the larger model still wins on rule induction, entailment and Lean formalisation. Every figure is webAI's, measured by webAI.
It has not been bought at the cost of general reasoning, but it has not gained much either. Across webAI's ten held-out general benchmarks it averages 0.7901 against 0.7942 for the base, behind Qwen3-8B (0.8493) and VibeThinker-3B (0.8097). It scores 95.4% on the BIG-Bench Hard logic subset.
The failure modes are specific. Answers on the logic tasks average 1,902 tokens and 24.2% hit the length cap, exact-match first-order-logic translation scores 0.01, and the scores were all taken with greedy decoding on BF16 weights in an 8,192-token window, while the shipped config samples by default and the GGUF files are untested. The 131,072-token context is inherited from Granite, not validated.
The licence decides most cases. The webAI Non-Commercial License limits the weights and any derivatives to non-commercial research and education, so a company that wants a local logic checker in a product needs a separate deal, and none is priced. For a researcher who wants a laptop-sized model to translate and check arguments, with a solver to verify the output as webAI advises, it is a sensible download.
Where it earns its place
A philosophy or computer-science lecturer
Check whether students' argument conclusions follow from their premises
Run the Q4_K_M GGUF in llama.cpp on a teaching laptop; entailment accuracy on webAI's tests rose to 67%. Educational use is within the licence.
A formal-methods research group
Critique candidate Lean statements before sending them to a prover
webAI reports 79.5% on Lean proof critique, up from 66% for the original TwIL-LM3. Use greedy decoding to match the reported scores and verify output with the prover.
A privacy-focused research lab
Turn natural-language rules into structured logic without sending text to a cloud API
The whole model fits in 2 to 7 GiB depending on quantisation and runs offline. Exact-match first-order-logic translation scores only 0.01, so treat output as a draft for a symbolic solver.
Before you commit
- Setup
- Download from Hugging Face; Q4_K_M GGUF in llama.cpp for CPU or a 4 GB GPU, BF16 in Transformers, Q4_0 for GenieX on Snapdragon NPUs with the context raised from its 4,096-token default.
- Integrations
- Transformers (GraniteForCausalLM), llama.cpp GGUF builds, a LiteRT-LM build for on-device runtimes, GenieX.
- Data and privacy
- Runs locally, so nothing leaves your machine. The webAI Non-Commercial License 1.0 limits the weights and derivatives to non-commercial research and education; IBM's Apache 2.0 base licence is kept in the repo. No safety tuning beyond Granite's.
- Cost at scale
- Free to run on hardware you own; the Q4_K_M file is 2.09 GiB and F16 is 6.82 GiB. webAI estimates about 792 output tokens per answer on its server tests. Commercial use has no published price.
- Limits
- Validated only at 8,192 tokens of context; English only; not tested on code or tool use; procedural reasoning scores 0.12; not a general assistant.
What’s good
- Matches Qwen3-8B on webAI's formal-logic composite at 3.66B parameters
- Q4_K_M GGUF is 2.09 GiB and runs on a CPU or a 4 GB GPU
- Formal-logic gains without losing general benchmark average
- Detailed model card with failure rates and evaluation settings
What’s not
- Non-commercial licence; no commercial terms published
- Verbose: 24.2% of logic answers hit the length cap
- Behind Qwen3-8B and VibeThinker-3B on general benchmarks
- All benchmarks self-reported; quantised builds untested
Best for
Researchers and students who want a local model for entailment checks, first-order-logic translation and Lean critique, and do not need commercial rights.
TwIL-LM3-Pro alternatives
TwIL-LM3-Pro: common questions
- Is TwIL-LM3-Pro worth it?
- A small model that does one kind of logic well enough to match models twice its size, under a licence that keeps it in the lab. We rate it 3.5 out of 5.
- How much does TwIL-LM3-Pro cost?
- Free weights for non-commercial research and education under the webAI Non-Commercial License 1.0. No hosted API or commercial licence price published.
- What are the downsides of TwIL-LM3-Pro?
- Non-commercial licence; no commercial terms published; verbose: 24.2% of logic answers hit the length cap; behind Qwen3-8B and VibeThinker-3B on general benchmarks; all benchmarks self-reported; quantised builds untested.
- Who is TwIL-LM3-Pro best for?
- Researchers and students who want a local model for entailment checks, first-order-logic translation and Lean critique, and do not need commercial rights.
- How much does TwIL-LM3-Pro cost for a team or in production?
- Free to run on hardware you own; the Q4_K_M file is 2.09 GiB and F16 is 6.82 GiB. webAI estimates about 792 output tokens per answer on its server tests. Commercial use has no published price.
- Is TwIL-LM3-Pro safe for company data?
- Runs locally, so nothing leaves your machine. The webAI Non-Commercial License 1.0 limits the weights and derivatives to non-commercial research and education; IBM's Apache 2.0 base licence is kept in the repo. No safety tuning beyond Granite's.
- What are the best TwIL-LM3-Pro alternatives?
- The closest alternatives we have reviewed are VibeThinker-3B, Granite 4.2 and Qwen3.8-27B.
Answers checked Nov 11, 2026.