Torad Labs · Research
Measured · re-runnable
Research is the product.
Open methods, re-runnable benchmarks, and the numbers behind every claim on this site.
Every number here is re-runnable.
Each result names the model it ran on and the method behind it. Open items live on /technology.
0.9998similarity vs bf16 original
31%of vLLM memory
238 tok/svs 267 on vLLM
5 / 5top answers matched
What we have measured
VALIDATED| What we claim | What we measured |
|---|---|
| It needs about a third of the memory | 870 MB vs 2.7 GB (31%). Qwen3-1.7B linear weights, Torad Quant 4-bit, our own format. It stores each number using 4 bits instead of the usual 16, so the model takes about a third of the space in practice. against Short for bfloat16, the standard 16-bit format AI models normally run in. Accurate, but it takes a lot of memory.. |
| It writes about as fast as the cloud software | 238 vs 267 How fast the model writes. A token is roughly a short word or a few letters, so this is close to words per second., within 10%. Qwen3-1.7B against The most popular open-source software for running large AI models. Fast, but it needs a lot of memory to load a model.. |
| It runs models the cloud software cannot even fit | 40 and 21 tokens per second for Qwen3-8B and Qwen3-14B (8 and 14 billion The adjustable values inside a model. More parameters usually means a bigger, more capable, and more memory-hungry model. An "8B" model has 8 billion of them.) on an ordinary 16 GB graphics card, where vLLM cannot load them at all. |
| The shrunk model gives the same answers as the original | 0.9998 A score from 0 to 1.00 for how alike two answers are, called cosine similarity. 1.00 means identical. We use it to compare the shrunk model against the full-size original. on Qwen3-1.7B, top answer matching on 5 of 5 test prompts (The shrunk model gives the same answers as the full-size original, down to the same top choice.). |
| It works on other model families too | 0.9997 on 4 of 5 for Gemma 4 E2B, a different base model, the same way. |
| It takes very little code to run | 2,500 lines of Rust and 1,300 of CUDA, against roughly 500,000 in the standard stack. |
Posts on training, shrinking, and geometry.
In the order they were written. These links open blog.torad.ai.
- TrainingSFT copies, GRPO discoversFine-tuning imitates. Reinforcement finds. Why the split shapes the whole pipeline.
- MethodRejection architectureBuild what a model refuses into its design, instead of patching it in later.
- TheoryImpedance matching in communicationTransfer peaks when source and load match. Applied to models and the people who use them.
- MethodManufactured discoveryHow to make a model find the answer instead of being told it.
- The labIntroducing EliThe orchestrator behind the lab.
Three preprints under way.
They land here as they finish.
- A 4-bit model that can still learn ASSERTED in progress How we keep a shrunk 4-bit model trainable. Most methods freeze it for good. No one else has published this.
- Running and training in one step ASSERTED in progress The megakernel: running a model and training it in one pass, instead of two separate programs.
- Shrinking a model without losing it ASSERTED in progress How TQ4 climbs from the naive baseline (0.85) to the same answers as the original (0.9998). The full write-up of the result above.
How claims are labeled.
VALIDATED means measured and re-runnable: a named model, a stated method, a number you can reproduce.
ASSERTED means believed but not yet proven, like the drafts above.