8 GB / RTX 3060 Ti
49.71 tokens/s
6,118 MiB of graphics memory used by the test process with 131,072 text tokens already in context.
Preliminary research measurement ·
Research / In progress
Measured,
with clear limits.
Our goal: AI that can work with more of your documents on a graphics card you can own.
Tests across consumer graphics cards
A selection from our September 2026 hardware tests, grouped by the memory available on each card. All figures below use the same workload: one conversation with 131,072 text tokens already in context, generating 64 more tokens.
12 GB / RTX 3060
39.98 tokens/s
6,084 MiB of graphics memory used by the test process with 131,072 text tokens already in context.
Preliminary research measurement ·
12 GB / RTX 4070 Ti
82.89 tokens/s
6,128 MiB of graphics memory used by the test process with 131,072 text tokens already in context. This card was tested locally; the others shown were rented.
Preliminary research measurement ·
16 GB / RTX 4080
104.02 tokens/s
6,231 MiB of graphics memory used by the test process with 131,072 text tokens already in context.
Preliminary research measurement ·
24 GB / RTX 4090
140.01 tokens/s
6,397 MiB of graphics memory used by the test process with 131,072 text tokens already in context.
Preliminary research measurement ·
What these measurements show
These tests used random model weights: they measure running speed and memory use, not useful answers. Tokens are pieces of text. Speed here counts new tokens after the input has been processed; it is not the time from asking a question to receiving an answer. Each test ran one conversation. Results describe these test setups, not a promise for your computer. There is no released model or public AI service here.
Memory figures count only the test process, measured after start-up minus memory use before start-up. They exclude desktop and other programs. They are not a measurement of answer quality or proof that the system can correctly use all the text.
The test used the same model file layout with random values in place of trained weights. It ran our llama.cpp fork at build b671076fb, with Q6_K shared weights, MXFP4 expert weights and a Q4_0 text-memory cache. These technical settings identify the test; they are not requirements for a released product. The local card used Windows and CUDA 13.4; rented cards used Linux and CUDA 12.8.
Source: the committed hardware sweep dated September 27, 2026. These five examples are measured results, not averages for a memory class or a ranking of cards. Other cards and host conditions can give different results.
What remains unmeasured
This comparison provides no results for cards with 6 GB or less, laptop graphics, or AMD and Intel graphics. Larger memory allocations in the source do not demonstrate a usable million-token conversation. Answer quality needs separate tests with trained models.
Roadmap
This is the path we are working toward, with no promised release date. Each step depends on the results of the one before it.
Develop and test
Build a research model and check its answers against tests chosen in advance.
Test on everyday hardware
Check speed, memory use and correct operation on consumer graphics cards.
Try complete tasks
Test the experience from question to answer under defined conditions.
Consider wider access
Decide what can be offered after the evidence and release review support it.