Buy

Q4T / Q3T

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

Small, and faithful.※1

Quantisation is the technique that shrinks an AI model so it can run on a Mac. The models in R.E.V.I.S. v1.3.0 are built with our own quantisation recipes, Q4T and Q3T. The goal is simple: make the model smaller while keeping its answers close to the original.

4-BIT / 3-BITDATA-FREE

What is Quantization

Quantisation means using fewer digits for the weights.

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

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

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

Smaller

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

Faithful

Knowing what to trim and what to keep is what decides 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 according to the role of each part.

Most of the modelThe expert layers, etc.4 bits in Q4T, 3 bits in Q3T. Shrinking this part is what makes the whole model much smaller.
Parts that shape the answerKept fine-grained.
Parts that choose which experts to use, etc.Left untouched, just as in the original.These parts are delicate: even a tiny difference can swap which expert gets chosen.

Not Just Rounding

More than just cutting digits.

Standard quantisation simply rounds each weight, one by one, to the nearest step on a scale. With an image, that's like lowering the resolution into a mosaic. Q4T and Q3T transform the weights first and then pack them, much like JPEG. For the same amount of data, the information preserved is completely different.

Original image

Standard quantisationLower resolution (mosaic)

Q4T / Q3TTransform, then pack (like JPEG)

A visual comparison. Both the middle and right images use 1/36 of the original data. Because JPEG also includes adjustments tuned for the human eye, the exact mechanism isn't identical.

Standard quantisation

Round each weight to the nearest step

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

Q4T / Q3T — 1

Smooth out the spikes first

Just as JPEG transforms an image before packing it, the weights are rotated mathematically (Hadamard rotation) to spread outliers across the other values. With the range evened out, even a small number of bits can use finely spaced steps.

Q4T / Q3T — 2

Pick the closest path along each sequence

Instead of rounding weights one by one, we pack them together as sequences (trellis coding). From the possible paths, we choose and record the one with the smallest overall error against the original sequence (dotted line).
Standard quantisationQ4T / Q3T
How weights are roundedOne by one, to the nearest stepAlong each sequence, picking the path of least error
OutliersForce the steps to become coarserEvened out before packing
How different parts are treatedRoughly the same bit width throughoutSet by part (4 or 3 bits / kept finer / unchanged)

Data-Free

No calibration data used.

When you quantise a model to fit sample data (calibration data), you can show strong results in that domain, but there is no way to tell what has changed everywhere else. Q4T and Q3T use no calibration data at all. They are built solely through a mathematically defined process that minimises the error from the original weights.

Quantisation tuned to calibration dataQ4T / Q3T
Domains covered by the dataCan show strong resultsMeasured by closeness to the original model
Everything elseNo way to check what has changedMeasured by the same yardstick (closeness to the original)
How it's builtResults depend on the chosen dataA purely mathematical process

Not skewed to 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

There is no training that alters the model's internals. Every model is a direct quantisation of the publicly released original.

Verified by closeness

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

Measured

Smaller, yet closer to the original.

We had each model read the same texts and measured how much its choice of the next word 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, 4 to 6 times less drift

Compared with standard Q4 of roughly the same size, Q4T cut the drift to about a quarter on Qwen3.6 and about a sixth on Gemma 4.

Smaller and closer than standard Q5

Q4T is smaller than standard Q5, yet drifts less from 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 KLD (Kullback–Leibler divergence) of the next-word probability distribution against the original model (BF16), measured across 48 documents and roughly 18,000 positions. 'Standard' refers to models quantised by us using standard affine quantisation under favourable conditions (keeping output layers and similar parts at higher precision). Measured in a research environment by computing directly from the 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 provides 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.