Buy

Q4T / Q3T

Quantization by Q.E.L.V.A.R.I.

Small, and faithful.※1

Quantization is the technology that makes AI models small enough to run on a Mac. The models in R.E.V.I.S. v1.3.0 are built with our own quantization recipes, Q4T and Q3T. Our goal was to make them smaller while keeping their answers close to the original model.

4-BIT / 3-BITDATA-FREE

What is Quantization

Quantization means using fewer digits for the weights.

An AI model is made of billions of numbers called weights. The original model stores each weight in 16 bits. Cut that to 4 bits, and the model is roughly a quarter the size—making it much easier to fit into your Mac's memory. Cut it carelessly, though, and the quality of the answers drops.

Size of Qwen3.6 35B (weights)
  1. Original modelAbout 70 GB
  2. Q4T19.2 GB
  3. Q3T15.2 GB

Each square shows the bits per weight for most of the model (16 in the original, 4 in Q4T, 3 in Q3T).

Small

16-bit weights down to 4-bit or 3-bit. Even a 35B model runs on a Mac with 24GB of memory (Q3T).

Faithful

Where to trim and where to keep. That choice determines how close the answers stay to the original model.

On your Mac

The smaller model runs right on your Mac. Your conversations never need to be sent to the cloud.

The Recipe

Different bit widths for different parts.

Inside a model, some parts hold up well with fewer bits, while others are sensitive to the slightest error. Q4T and Q3T are fixed recipes that set the number of bits for each kind of part.

Most of the modelThe expert layers, etc.4 bits in Q4T, 3 bits in Q3T. Making this part smaller is what shrinks the model as a whole.
Parts that shape the answerKept fine-grained.
Parts that choose which experts to use, etc.Not shrunk—kept as they are in the original.Because this is a delicate part where a tiny difference can swap which expert gets chosen.

Not Just Rounding

More than just dropping digits.

Standard quantization simply rounds each weight, one by one, to the nearest step. In image terms, it's like lowering the resolution into a mosaic. Q4T and Q3T transform the weights first and then pack them, like JPEG. For the same amount of data, the information that survives is completely different.

Original image

Standard quantizationLower the resolution (mosaic)

Q4T / Q3TTransform, then pack (like JPEG)

A visual analogy. Both the middle and right images use 1/36 of the original data. Since JPEG includes tuning for the human eye, the exact mechanism isn't identical.

Standard quantization

Round each weight to the nearest step

Each weight is rounded to the nearest step on its own. If even a single outlier (far right) is present, the step size has to stretch to cover it, making all the other values coarser.

Q4T / Q3T — 1

Even out the spread first

Just as JPEG transforms an image before packing it, the weights are mathematically rotated (Hadamard rotation) to spread outliers across the other values. With a uniform range, the steps stay fine even with fewer bits.

Q4T / Q3T — 2

Pick the closest path along a sequence

Instead of rounding weights one by one, they are packed together as a sequence (trellis coding). Out of the possible paths, we pick and record the one with the smallest total error from the original sequence (dotted line).
Standard quantizationQ4T / Q3T
How weights are roundedOne by one, to the nearest stepAlong a sequence, picking the path with the least error
OutliersMake the steps coarserEvened out before packing
How parts are treatedRoughly the same bits everywhereSet per part (4 or 3 bits / kept finer / unchanged)

Data-Free

No calibration data.

When you quantize a model to fit sample data (calibration data), you can show strong numbers in that domain—but there's no way to tell what changed everywhere else. Q4T and Q3T use no calibration data. They are built purely by a mathematical procedure that minimizes error against the original weights.

Quantization tuned to calibration dataQ4T / Q3T
Domains it was tuned forCan show strong resultsShown by closeness to the original model
Everything elseNo way to verify what changedMeasured by the same yardstick (closeness to the original)
How it's builtResults depend on the data chosenA mathematical procedure alone

Not tilted toward any domain

We don't tune for specific topics or languages. The only goal is to stay close to the original model, however you use it.

No retraining

We don't do any training that alters what's inside. Each model is a direct quantization of the publicly released original.

Verified by closeness

That means quality can be measured with a single yardstick: how close it is to the original model. The charts below show the results measured with that yardstick.

Measured

Smaller, yet closer to the original.

We had each model read the same passages and measured how much its word choices drifted from the original model. Shorter bars mean answers closer to the original.※2

Qwen3.6 35B MoEDistance from the original model (shorter is closer)
  1. Standard Q837.7 GB0.0066
  2. Q5T23.8 GB0.0090
  3. Standard Q629.3 GB0.0116
  4. Q4T19.2 GB0.0181
  5. Standard Q524.9 GB0.0248
  6. Q3T15.2 GB0.0356
  7. Standard Q420.7 GB0.0682
Gemma 4 26B MoEDistance from the original model (shorter is closer)
  1. Standard Q828.0 GB0.0110
  2. Q5T19.4 GB0.0136
  3. Standard Q621.9 GB0.0268
  4. Q4T15.6 GB0.0373
  5. Standard Q518.6 GB0.0638
  6. Q3T12.5 GB0.0650
  7. Standard Q415.5 GB0.2367

At the same size, 1/4 to 1/6 the drift

Compared with standard Q4 of roughly the same size, Q4T's distance from the original was about 1/4 on Qwen3.6 and about 1/6 on Gemma 4.

Smaller and closer than standard Q5

Q4T is smaller than standard Q5, yet stays closer to the original. That holds true for both models.

Even at 3 bits, closer than standard Q4

Even the smaller Q3T stays closer to the original model than standard Q4.

Models

Models you can run.

ModelSize (Q4T / Q3T)
Gemma 4 26B MoE15.6 GB / 12.7 GB
Qwen3.6 35B MoE19.4 GB / 15.6 GB
Ornith 1.5 35B MoE19.4 GB / 15.6 GB
Qwen3.8 27B20.1 GB / 17.6 GB

Sizes include the MTP draft model (GiB). Models can be loaded within roughly your installed memory minus 7GB. As a guide, a Mac with 24GB of memory runs the Q3T of Gemma 4, Qwen3.6 and Ornith 1.5, and 32GB or more runs the Q4T of every model.

Back to Q.E.L.V.A.R.I.

We make it smaller
to keep what matters.

SEVEN CORES / ONE CONSCIOUSNESS

Seven brains. One mind.

  1. R.E.V.I.S.COGNITIVE OS
  2. Q.E.L.V.A.R.I.INFERENCE ENGINE
  3. F.A.L.Q.TASK QUEUE
  4. M.μ.X.I.S.TASK SCHEDULER
  5. J.I.T.R.X.PROMPT COMPILER
  6. L.A.T.R.I.O.RESPONSE RENDERER
  7. K.E.C.H.I.PERSONAL CONTEXT

Turn your Mac into real AI.

See editions and pricing Back to R.E.V.I.S.

※2 Average KL divergence (KLD) of next-token probability distributions against the original model (BF16), measured across 48 documents and roughly 18,000 positions. "Standard" models were quantized by us using standard affine quantization under favorable conditions (keeping output layers and similar parts at higher bit widths). Measured in a research environment by computing exact weight values. Sizes are weights only (GB). Q5T is the 5-bit version of the same recipe; R.E.V.I.S. v1.3.0 uses Q4T and Q3T.

※1 Accuracy refers to how faithfully the original model's output is reproduced; it does not mean performance beyond the original model. Speed varies with the model, the Mac and the input.