Introducing Leto 1Our first field-based AI model
Leto 1
- Published
- By
- Mumbrane
- Reading time
- 2 min
Leto 1 is our first model. From up to 3.25 million tokens of your data it builds a specialist model for your work, and it costs less to run than the big frontier models.
Leto 1 is the first AI model from Mumbrane. It doesn’t hand everyone the same general model. It builds a specialist model for each company, made from that company’s own data.
General models know a little about everything and look things up when you ask. Leto 1 reads your data once and turns it into what we call a field. The field holds what your data says and how the pieces connect, and that field is the specialist. It doesn’t search your files for an answer. It works each question out from what it already knows about your data.
Benchmarks against frontier models
We tested Leto 1 on two public benchmarks and put it next to the three highest-scoring models on each.
Legal reasoning across 162 tasks.
| Benchmark | Model | Score |
|---|---|---|
| LegalBench | Leto 1 | 92.44 |
| LegalBench | Claude Fable 5 | 88.56 |
| LegalBench | Claude Fable 5.1 | 88.51 |
| LegalBench | Gemini 4 Argon | 88.30 |
| τ³-bench | Leto 1 | 75.43 |
| τ³-bench | Qwen3.8 Max | 51.30 |
| τ³-bench | Grok 4.6 (high) | 50.70 |
| τ³-bench | Muse Spark 1.3 (max) | 50.50 |
Leto 1 came out ahead on both. A specialist doesn’t need to know everything, so it’s also much smaller than any of those models, which makes it faster and cheaper to run.
LegalBench covers 162 legal reasoning tasks. Leto 1 scored 92.44, ahead of Claude Fable 5 (88.56), Claude Fable 5.1 (88.51) and Gemini 4 Argon (88.30). In τ³-bench, a model handles banking requests with tools while talking to a simulated user. Leto 1 scored 75.43. The next best, Qwen3.8 Max, scored 51.30.
Hallucinations
In our tests, Leto 1 didn’t make anything up. Every answer it gave was backed by the data it had. When the data couldn’t support an answer, it told you so.
| Model | Hallucination rate |
|---|---|
| Leto 1 | 0.00% |
| GPT-6 Astra | 0.62% |
| Claude Opus 5.5 | 1.94% |
| Gemini 4 Argon | 2.37% |
| DeepSeek V4.1 | 3.81% |
| Grok 4.7 | 4.60% |
Each answer also points to the part of your data it’s based on. You can check it against the source in a few seconds.
A 3.25 million token context window
Leto 1 can take in up to 3.25 million tokens at once. That’s several thousand pages, more than any of the frontier models below.
| Model | Context window (tokens) |
|---|---|
| Leto 1 | 3,250,000 |
| GPT-6 Astra | 1,050,000 |
| Gemini 4 Argon | 1,048,576 |
| Claude Opus 5.5 | 1,000,000 |
| DeepSeek V4.1 | 1,000,000 |
| Grok 4.7 | 500,000 |
Your data is built into the specialist once, so a question never has to reread it. The tenth question on the same data takes about as long as the first, no matter how much data you’ve given it.
We tried going bigger. At 10 million tokens the model still answered, but its answers got worse on material far from the question, and nothing in the output warned you. 3.25 million was the largest size that held up in our tests, so that’s what we’re shipping.
Access through the Hailmary program
Companies in the Hailmary program get Leto 1 first. It’s a small group running the model on their own data, with our team working alongside them. If your company wants in, you can apply here. After that, access opens to people on the waitlist.