From a browser tab to a GPU cluster. Pick the one that fits you, then start a guided route or copy the official command.
Six sizes, from small devices to GPU servers. On a laptop, the guide picks a size for you and the check level confirms it. The server sizes run with vLLM or SGLang, or through a hosted API.
| Size | Made for | Good at | Watch out | Download | Speed* |
|---|---|---|---|---|---|
| On a laptop | |||||
| 0.9B | 8 GB computers, quick commands | Instant replies, reminders, simple tool calls | Slips on reasoning. Not for coding. | 0.7–1.2 GB | ~165 tok/s |
| 3.7B | 16 GB laptops | Summaries, document Q&A, tool calls, code fixes | Arabic phrasing was partly wrong in testing | 3.2–5.4 GB | ~46 tok/s |
| 7.0B | 24 GB+ laptops and workstations | Passed every check in our tests | Slowest of the three. Needs about 16 GB free. | 5.6–9.6 GB | ~40 tok/s |
| On GPU servers | |||||
| 32B | GPU servers (recipe validated on 2× H200) | Large dense model with a 512K context window | Too big for laptops | BF16 · FP8 · NVFP4 | — |
| 36B MoVA-36B-A4B | GPU servers (recipe validated on 2× H200) | Mixture of experts: 36B stored, only 4B active per token | Needs expert parallel in vLLM or SGLang | BF16 · FP8 | — |
| 375B 375B-A23B | One 8× H200 node | The flagship. Agentic tool use, terminal and long multi-step work | Needs a full 8-GPU node | BF16 · FP8 · NVFP4 | — |
*Measured on an Apple M4 Pro Mac with 24 GB. A token is about ¾ of a word. Computers without a GPU are several times slower. For server speeds, see the vLLM and SGLang recipes. All models are on Hugging Face.
Each of these happened during real testing. Search for what you see, or paste your error into any level's check box and the guide spots it for you.
No match. Ask in the IFM Discord at ifm.ai/discord and paste the last lines of server.log.