An inference engine built from scratch for the Mac. Custom kernels optimized for Apple silicon push speed to the limit. Our own quantization runs models light while keeping accuracy close to the original. Inference happens on your Mac, and conversations are never sent to the cloud.
Many local AI apps run models through a Python runtime or a local server. Q.E.L.V.A.R.I. is a native engine written only for the Mac, and it runs models directly inside R.E.V.I.S.
Built for Apple silicon
Built on MLX, with custom Metal kernels that tighten the computation. It turns the design of Apple silicon, where the CPU and GPU share the same memory, directly into speed.
Light and faithful
Our own quantization (Q4T, Q3T) makes models smaller and lighter while keeping accuracy close to the original model's output.※1
There's no network traffic for inference. Your conversations, your files, and the context the model has read all stay on your Mac.
Speculative Decoding
Draft ahead. Check in one pass.
With MTP (Multi-Token Prediction), a small draft model writes a few words ahead, and the main model checks them all at once. Only words the main model has confirmed are kept. The answer moves forward in fewer steps than writing one word at a time.
When the draft misses, the main model rewrites it. The speedup varies with the model, the context length, and how many run at once.
DECODE / MTP
DRAFT
VERIFY
OUTPUT
I've put the decisions and owners from the meeting minutes into a table.
Context Memory
Continue without rereading.
The context the model has finished reading (the KV cache) is encrypted and kept on your Mac's SSD. When you resume a conversation, it doesn't have to read everything again from the start.
Encrypted, kept local
Saved context is encrypted, and the next turn of the conversation picks up from there.
Several at once
Conversation replies and generation for background tasks can run side by side on a single Mac.
Semantic search, in the same place
The same engine also runs the embedding models that turn the meaning of words into vectors. It's the part that supports K.E.C.H.I. search.
Measured
By the numbers.
It doesn't reread everything from the start on every turn of a conversation. It keeps the context it has finished reading on hand, reads only what was just added—your latest message or a tool's result—and writes the answer. So at each length of context already read, we measured how fast it reads new input and how fast it writes the answer across five Macs. Every one of these speeds was achieved without sacrificing computational precision.※2
How Macs compare.
We compared speeds using Qwen3.6 35B and Gemma 4 26B, which run on every Mac. These are the speeds when the context the model has already read is 16K (a typical length in everyday use).
Qwen3.6 35BGemma 4 26B
Prefill (how fast it reads new input)tok/s
M5 Ultra1,4431,505
M5 Max1,2381,284
M4 Max664628
M5 Pro783815
M4 Pro390372
Decode (how fast it writes the answer)tok/s
M5 Ultra239241
M5 Max196170
M4 Max172144
M5 Pro123106
M4 Pro10386
A feel for writing speed.
This replays decode (how fast it writes the answer) at real speed. It shows the speed for Qwen3.6 35B with 16K of context. Choose a Mac and press Play.
0.0 s
People read silently at about 4 words a second.
0.757words/s230
Expense claims go through three steps: submission, approval, and payment.
First is submission. Within one month of the expense, enter the date, amount, and purpose in the claim form, attach a photo of the receipt, and send it in. For transit fares, you don't need a receipt if you list the route and fare. However, receipts are required for taxis and bullet trains.
Next is approval. Once your claim arrives, your manager checks the details and approves it. Claims over 50,000 yen also need approval from the department head. Incomplete claims are sent back, so it's best to state the purpose specifically—including who it was with and what it was for.
Last is payment. After accounting verifies the claim, anything approved by the 20th of the month is deposited with your paycheck on the 25th. Claims approved after the 20th are paid the following month.
If you're unsure about anything, check "How Expense Claims Work" on the company portal first. If you still have questions, contact the accounting desk—they usually reply the same day.
Context length and speed.
As the context the model has already read gets longer, both reading speed and writing speed go down. The horizontal axis is context length (1K–128K tokens), and the vertical axis is speed. Choose a Mac.
Qwen3.6 35BOrnith 1.5 35BGemma 4 26BQwen3.8 27B
Prefill (how fast it reads new input)tok/sDecode (how fast it writes the answer)tok/s
Prefill (how fast it reads new input)tok/sDecode (how fast it writes the answer)tok/s
Prefill (how fast it reads new input)tok/sDecode (how fast it writes the answer)tok/s
Prefill (how fast it reads new input)tok/sDecode (how fast it writes the answer)tok/s
Prefill (how fast it reads new input)tok/sDecode (how fast it writes the answer)tok/s
※2 Measured with the same settings as the R.E.V.I.S. release build. Speed varies by model, Mac, and input. tok/s is the number of tokens (pieces of words) processed per second.
Models
Models you can run.
Only models confirmed to work with Q.E.L.V.A.R.I. can be loaded. Support for new models will be added over time.
Model
Size (Q4T / Q3T)
Gemma 4 26B MoE
15.6 GB / 12.7 GB
Qwen3.6 35B MoE
19.4 GB / 15.6 GB
Ornith 1.5 35B MoE
19.4 GB / 15.6 GB
Qwen3.8 27B
20.1 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.
※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.
※ The animations on this page are illustrations and differ from the actual product screens.