Project · Research · In validation
Engine-Aware Quantization, built for the engine that runs it.
EAQ is not just another quant. It is a quantization framework designed around how TensorFold actually runs models, tuned for coding, agentic work and tool calling. The first target is GLM 5.3 Flash, and the early results are in.
The idea
A quant should know what will run it.
Quantization makes a big model small enough to run at home, and every quant gives up a little of the original model to get there. The question is where that loss lands.
A quant that is made with no idea of the engine behind it can lose quality exactly where it hurts. EAQ starts from the other end: from how TensorFold runs the model, and from what the model has to get right when it is writing code or driving an agent.
Designed around the engine
Each EAQ quant is made for how TensorFold actually loads and runs the model, not as a generic file that any engine might pick up.
Measured on the output
Quality is judged against the original BF16 model on what the quant actually writes, and on the agent decisions it makes along the way.
A framework, not a one-off
The same method is meant to carry over to every model and every quant I build from here on, not just this first one.
Early results
Closer to the original, at the same speed.
EAQ against my current public quant of GLM 5.3 Flash. Same engine, TensorFold, and both measured against the original BF16 model.
Early numbers from my own runs. EAQ is still being validated on real tool-calling and agent benchmarks, and the full details come with the release.
Built for agents
The decisions that make or break a run.
In coding and agent work, a single wrong decision can derail a whole task, even when every sentence reads fine. That is why EAQ is measured on these three, not only on the text.
When to call a tool
Answer from what it knows, or reach for a tool? Calling at the wrong moment sends an agent loop off course.
Which tool to call
Reading a file is not running it. Picking the wrong tool costs a step, or does the wrong thing.
When to stop
Stop too early and the job is half done. Keep going and it burns time and tokens on work nobody asked for.
Where it stands
From the first model to every model.
- Done
Built
The EAQ framework, designed around TensorFold.
- Done
First target
GLM 5.3 Flash, with early results against my current public quant.
- Now
Validating
Real tool-calling and agent benchmarks, before anything goes public.
- Next
Release
The first EAQ quant, with the full details and numbers.
- Next
Every model after
The framework carries over to the models and quants I build next.
Questions
Before you ask.
Can I download it?
Not yet. EAQ is still being validated on real tool-calling and agent benchmarks, and nothing goes public before that is done. Follow me on X to hear the moment it ships.
What is it compared against?
My current public GLM 5.3 Flash quant, on the same engine, TensorFold. Both are measured against the original BF16 model, so the numbers show how much closer EAQ stays to the full-precision model.
Is it slower?
In the early runs, roughly the same speed as the current quant.
Will it come to other models?
That is the point of making it a framework: it is meant to carry over to every model and quant I build in the future. GLM 5.3 Flash is the first.
Can I use it with another engine?
EAQ is built around how TensorFold runs models, and the results above were measured on TensorFold. It is the engine behind my recipes today.
More details coming soon.
Follow along on X for the release. Until then, my TensorFold recipes run the current quants.