memown FAQ
What we measured, when, and on which hardware, and what memown does not do. Every figure on this page is an internal measurement by ZMLabs on real hardware, and each one is dated.
Measured figures
Can memown run a model bigger than the graphics card's memory?
In an internal test in June 2026, with a version of memown before 1.0, yes. A 120B-class model (gpt-oss-120B, about 63 GB) started and served on a laptop with one 8 GB graphics card (RTX 4070 Laptop) and 16 GB of RAM, under Windows 11 (4 successful starts), reading most of the model from its SSD. On the same laptop, llama.cpp alone (the engine memown uses) generated about 1.5 tokens per second: it runs, slowly. The speed comes from a single run. This has not been re-measured with version 1.0.
How fast is a 30B model on an 8 GB laptop graphics card?
A 30B model (Qwen3-30B-A3B, 4-bit) on an RTX 4070 Laptop graphics card with 8 GB of memory: about 14 to 21 tokens per second at a safe memory setting, 22 at best at a setting close to the memory limit. Measured in June 2026, internally, with llama.cpp, before version 1.0.
How much sooner does the answer start when a saved context is reused?
Reusing a saved context instead of re-reading it made the first token of the answer arrive 4.0 times sooner for a 179-token context and 14.7 times sooner for a 2,048-token context (Qwen3-30B-A3B on a rented datacenter graphics card, NVIDIA L40S 48 GB). This tested llama.cpp's save-and-restore, the mechanism memown uses, directly with llama.cpp, with 50 runs per length. Measured on July 3, 2026, internally.
What does version 1.0 do after a restart?
After memown is restarted, a saved context is reused without being re-read; only the new question is processed. Measured on September 23, 2026 on Windows 11 and Ubuntu 24.04: 214 saved tokens restored, 29 new tokens computed.
Is the answer the same after a restart?
On the same machine, yes: an answer computed from a saved context after a restart is identical, byte for byte, to the same answer computed from scratch (5 of 5 answers identical on Windows 11 and on Ubuntu 24.04; small model of 398 MB, without a graphics card). Measured on September 23, 2026.
Honest limits
Does memown make a small computer fast?
No. When a model is much bigger than the graphics card's memory, answers are slow. With llama.cpp alone (the engine memown uses), the 120B model generated about 1.5 tokens per second on the 8 GB laptop with 16 GB of RAM, and 7.6 tokens per second on a rented server with 188 GB of RAM (NVIDIA L40S 48 GB). Measured in June 2026, internally.
Is the answer identical on Windows and on Linux?
No. Byte-identical holds on the same machine only: the same question gave different answers on Windows and on Linux (measured on September 23, 2026). With version 1.0, it has not been measured with large models or on a graphics card.
Does memown save energy?
We make no energy or environmental claim: we have not measured it.
Using memown 1.0
Does anything leave my computer?
After installation, nothing leaves your machine (by design). On first launch, memown downloads its engine (llama.cpp, from GitHub); you provide the model file (.gguf), for example downloaded from Hugging Face. After that, everything runs on your computer.
Which systems does memown 1.0 run on?
Windows (x64): verified on September 23, 2026 on a fresh Windows 11 machine without a graphics card; the program is not code-signed. Linux (x64): Ubuntu 24.04 or newer (glibc ≥ 2.38), with libgomp1 installed; verified on Ubuntu 24.04 on September 23, 2026; it does not start on Ubuntu 22.04. macOS (Apple Silicon and Intel): July 2026 archives, signed and notarized, not re-verified since.
How do apps talk to memown?
Through a local API: an OpenAI-compatible chat endpoint at http://127.0.0.1:8095/v1, which listens only on your own computer (it answered on September 23, 2026 on Windows 11 and Ubuntu 24.04). Start it with memown serve --model <file.gguf> (on Windows: .\memown.exe serve --model <file.gguf>).
Who measured these figures?
ZMLabs, internally, on real hardware; each figure is dated. The raw measurement files are not published.