Introducing Leto 1Our first field-based AI model
Leto 1
- Published
- By
- Mumbrane
- Reading time
- 3 min
Our first model. Faster and cheaper than the frontier models, it reads up to 3.25 million tokens of your documents and answers only from them.
Leto 1 is the first model from Mumbrane. It is built for one kind of work: answering questions from a large set of documents you supply, and showing what each answer rests on.
Most language models reread a long context every time you ask something. Leto 1 reads it once. It turns your documents, facts and definitions into what we call a field, a fixed structure that holds what the documents say and how the parts relate. Each question is then worked out inside that field.
Benchmarks
We ran Leto 1 on three public benchmarks and compared it with one strong model on each.
Legal reasoning across 162 tasks.
| Benchmark | Leto 1 | Rival | Rival's score |
|---|---|---|---|
| LegalBench | 72.44 | Fable 5 | 88.56 |
| τ²-bench | 94.60 | GLM 5.2 | 99.20 |
| ProcessBench | 82.61 | o1-mini | 87.90 |
Leto 1 scores a few points below each of them, and that is a trade we made on purpose. It is a much smaller model, so it runs faster and costs less than any of the three. A larger model wins points on broad benchmarks like these.
On LegalBench it gets 72.44 to Fable 5’s 88.56; many of its tasks are a short passage and a yes-or-no question, where size counts most. On τ²-bench, tool use in a conversation with a simulated user, it scores 94.60 to GLM 5.2’s 99.20. On ProcessBench, finding the first wrong step in a maths solution, it scores 82.61 to o1-mini’s 87.90.
Hallucinations
What we did not trade away is accuracy on your documents. In our tests Leto 1 made nothing up: every answer it gave was supported by the documents it had, and when they didn’t support one, it said so.
| Model | Maker | Hallucination rate |
|---|---|---|
| Leto 1 | Mumbrane | 0.00% |
| GPT | OpenAI | 0.62% |
| Claude | Anthropic | 1.94% |
| Gemini | 2.37% | |
| DeepSeek | DeepSeek | 3.81% |
| Grok | xAI | 4.60% |
Context window
Leto 1 takes in up to 3.25 million tokens at once, several thousand pages of text.
| Model | Maker | Context window (tokens) |
|---|---|---|
| Leto 1 | Mumbrane | 3,250,000 |
| Gemini | 2,000,000 | |
| Grok | xAI | 2,000,000 |
| Claude | Anthropic | 1,000,000 |
| GPT | OpenAI | 1,000,000 |
| DeepSeek | DeepSeek | 128,000 |
Because the documents are compiled once, a question doesn’t reread them. Later questions on the same documents take about as long as the first, however large the set is.
We tried larger sizes. At 10 million tokens the model still answered, but its answers got worse on material far from the question, and nothing in the output showed it. 3.25 million is the largest size that held up across our tests, so that is the limit we ship.
How it works
The field is an energy function over the model’s internal state. Each passage you provide adds a term to it, weighted by how much support it carries. To answer a question, the model starts from the question and moves the state toward lower energy, step by step, until it stops changing.
This is close to what attention already does. One attention step can be read as one update of a modern Hopfield network (Ramsauer and others, 2020). Leto 1 keeps taking those steps until the state settles, instead of taking one per layer.
Working this way gives two things. Each step is a weighted pull from specific passages, so the model can return the passages that shaped its answer. And when the state settles somewhere the passages support poorly, the model can tell, and says it can’t answer instead of guessing.
Limits
Leto 1 is a first version. The field stays fixed while a question is answered, so it can’t take in new information partway through. It is smaller than the models it was compared with, which costs it points on short, general tasks. And it loses accuracy on material far from the question as it nears 3.25 million tokens. These are what we are working on next.
The name
In Greek myth, Leto was turned away from every land until Delos, a floating island, stopped drifting and let her come ashore. It seemed right for a model whose answers come from settling.
Access
Leto 1 opens first to people on the waitlist, in small groups.