Speed · Trida2.0-4B
Faster text, same answer
2.59tokens / pass
Up to 3.10 on GSM8K. Every token is checked by the model's own autoregressive path, so quality matches it.
Concurrency 1, threshold 0.9, reasoning on, SGLang.
Same answers, faster. Decisions you can measure.
Trida is Trillion Labs' family of open models. Trida models write several words at once and check each one as they go, so you get the same answer as a standard model, only faster. TL-Decision answers yes/no, pick-one and rating questions with a probability for every option.
See the models
Open weights · Hugging Face
Write several tokens per pass
Diffuse
Trida models are language models that write several tokens per forward pass and still give the answer a normal model would. Use them like any chat model behind an OpenAI-compatible API.
InA prompt or a chat, with tool calls
OutText, about 2.6 tokens per forward pass instead of one
- Trida2.0-4B Open weights The Trida model to use today. Tuned for tool use and Korean agent tasks.
- Trida Stack Open source Convert your own model (Qwen3, Qwen3.5) into a block-diffusion model and serve it.
- GravityOCR ↗ Open weights Trillion Labs' document-OCR model, not a Trida model, drafting with the same technique: 9.7 tokens per pass on one H100 with the same output. It has its own site and Gravity for Mac app.
Make decisions
Decide
Ask typed questions about a piece of text and get a calibrated probability for every allowed answer, in a single forward pass. Software can branch on the answer directly.
InText (a message, a document, an agent log) and your questions: yes/no, pick one, or a score
OutA probability for every allowed answer. Nothing outside the options.
- TL-Decision-4B Open weights Decisions only: it never writes text, so it can't answer outside the options. Built on Trida2.0-4B; serve it with kev, same request format as Jev's System One API.
- Trida2.1-4B Coming soon Decisions and text in one model: a Trida model with typed decisions built in.
TL-Decision-4B · hard-v1 · 1,083 questions
Decisions only
0.867accuracy
The same backbone answering through a prompt scores 0.476.
English development split, scored with kev's harness.
Trida2.1-4B · Korean typed-decisions test · 2,000 questions
Decisions and text
0.762accuracy
TL-Decision-4B scores 0.654 on the same test split.
Test items screened for overlap with training data. All results and caveats →
How it works.
One model, two jobs.
How the Trida models and GravityOCR write faster: one model drafts with diffusion and verifies with autoregression, in the same forward pass. TL-Decision-4B reuses Trida2.0-4B's weights to answer in a single pass.
Fill a whole block
The diffusion stream predicts every masked token in the next block at once, instead of one token per forward pass.
Check it with AR
The same model's autoregressive path scores the draft. No second model, no extra weights.
Keep the AR answer
Matching tokens are kept and the first mismatch is replaced by the AR token, so the output equals plain autoregressive decoding.
Trida models and GravityOCR
Draft a block, verify it, commit
2 forward passes
One token at a time
Standard autoregressive decoding
15 forward passes
Get started.
Pick the shortest path to a first result.
App · no setup
Try Gravity for MacFree, offline OCR on Apple Silicon. Nothing is uploaded.
Model · one GPU
Serve TL-Decision-4B with kevA drop-in endpoint for System One-style decisions.
Notebook
Run a Trida model, step by stepServe Trida2.0-4B and compare AR with self-speculative decoding.
Notebook
Convert your own modelFirst-run block-diffusion training from an AR checkpoint.
Model · transformers
Run GravityOCR in PythonLoads with stock transformers; self-speculative serving via SGLang.
Why it matters
More work from
the same GPUs.
Every forward pass costs the same whether it commits one token or nine. Committing several verified tokens per pass is a direct way to get more out of an AI factory, and it is part of Trillion Labs' research on AI infrastructure that gets cheaper to run.
About Trillion Labs ↗A few details
What is block diffusion?
A language model that generates text block by block: autoregressive across blocks, but filling all positions inside a block in parallel, the way diffusion models fill in an image. Trida models keep a trained AR path in the same weights, which is what lets a model verify its own drafts.
Is the output really the same as autoregressive decoding?
Yes, at matched sampling settings. Every committed token is either a draft token that equals the AR token or the AR correction itself. The gain is tokens per forward pass, not a different answer. Under bf16 serving kernels, rare floating-point tie-breaks can differ.
How do the Trida models, TL-Decision and GravityOCR relate?
Trida is the family: the Trida models, TL-Decision and Trida Stack. The Trida models are the language-model line. TL-Decision is a Trida2.0-4B backbone fine-tuned to make typed decisions instead of writing text, and Trida2.1-4B builds those decisions into a model that also writes. GravityOCR is not a Trida model: it applies the same draft-and-verify method to document OCR, and Gravity for Mac is the app built on it.
What are the licenses?
Each model card states its license. Trida2.0-4B, TL-Decision-4B and Trida Stack are Apache-2.0; GravityOCR is MIT, inheriting GLM-OCR.