The model you already run.
More words a second.
Qwen2.5 7B, Q4_K_M on an RTX 3050 6 GB laptop. How it was measured
What VEYMEER is
VEYMEER serves open language models on the GPU you already own. On a 6 GB laptop GPU it writes 1.4 times as fast as llama.cpp at its best setting and 2.5 times as fast as Ollama, and more than four times as fast when it edits code. At temperature 0 the answer is the one the model gives on its own, word for word.
The other engine, at 1×. Qwen2.5 7B, Q4_K_M, RTX 3050 6 GB laptop.
Watch the same answer arrive
Four engines, one model file, one GPU, 256 tokens. This replays each at the speed we measured for it, in real time.
An engine built for one machine and the person in front of it
VEYMEER is built first for a single GPU answering a single person, and scaled up from there to several GPUs, several machines and a team.
The model's own answers
Nothing is approximated to get the speed. At temperature 0 an answer is the one the model gives decoding one word at a time; every release is checked against that.
Faster where you spend your time
Chat, explanations and new code come a third to a half faster than llama.cpp at its best. Editing, quoting and rewriting, where most of the answer is already on the page, come several times as fast.
JSON that always fits, sooner
An answer held to a JSON schema, a grammar or a tool call always has that form, and arrives faster than free text does, not slower.
Models larger than your memory
A model that does not fit the GPU still runs on it: what does not fit waits in system memory, and beyond that on disk. Nothing to configure.
Drop-in for what you use
The OpenAI API, and GGUF files of models from Llama and Qwen to Gemma 4 and gpt-oss, including the ones Ollama already downloaded. Point your client at 127.0.0.1:8080/v1 and change nothing else. The models →
Yours, on your machine
One program with nothing else to install. Models, prompts and answers never leave the machine; the engine does not report what you run or how much.
Where it wins, where it is level, where it is behind
Every number on this site comes from a run we can repeat: the same model file, the same prompts, the same GPU, each engine alone on it.
Writing speed, in every kind of work
Against llama.cpp with the same draft model: 1.2 to 4.4 times as fast, 1.4 overall. Against Ollama: 1.9 to 9.8 times, 2.5 overall.
Plain chat with no draft model
With nothing to write ahead from, both engines are limited by the same GPU: 34.4 tokens a second against llama.cpp's 33.7.
The first word on a new prompt
0.89 seconds against llama.cpp's 0.55, the price of exact arithmetic. A prompt seen before, or read ahead, starts in about a tenth of a second.
Every device in the machine, and the machines next to it
NVIDIA GPUs run the engine's own kernels and need only the driver. AMD, Intel and Apple GPUs run through the system's graphics interface. With no GPU, the processor does the work.
Several GPUs share one model. Several machines on the same network share one too: each lends what it has, and one of them answers.
What runs where, and what we have measured →The GPU still does the heavy work. What does not fit is brought to it as each pass needs it. Try your own machine
An editor, a chat and a coding agent on your own models
STRADA is the open-source editor built on the engine: an agent that reads, writes and runs code in your project, inline completions, and a chat that knows the files you have open.
It is free for everyone, companies included. Its installer brings the engine along the way a game brings its runtime.
About STRADA →Every tier runs at the same speed
They differ in scale: how many conversations at once, how large a cluster, whether the engine serves other machines. One person on one machine never needs to pay.
Individual
One person, one conversation at a time.
- The whole engine at full speed
- Every device the engine supports
- The OpenAI-compatible API on this machine
- 1 conversation at a time
- A cluster of up to 3 other machines
Individual-PRO
Several tools and windows sharing one engine.
- Everything in Individual
- 4 conversations at a time
- Any installed model a request names
- A cluster of up to 4 other machines
Individual-MAX
A workstation and the machines around it.
- Everything in Individual-PRO
- 8 conversations at a time
- A cluster of up to 10 other machines
Enterprise-MAX
Serving a team or a product from your own hardware.
- Everything in Individual-MAX
- Any number of conversations at a time
- Clusters of any size
- Serving other machines over the network
- Running as a system service
The first public release is being prepared
The engine and STRADA are built and measured on Windows. See what the release will contain, and how installing it works.