Claim vs measured
Repos & live GitHub stats
Sneak peek
Outputs & screenshots
Who it's for
Test plan
Rerun bundleYou can now run the biggest open AI models on a 4 GB graphics card. AirLLM loads a model one layer at a time instead of all at once, so it fits in a tiny amount of video memory: → Kimi K3, 2.8 trillion parameters, in 3.72 GB → DeepSeek-V3, 671 billion parameters, in about 12Open on X → · 68,819 views
local-models/runtimes claim tier
AirLLM
TESTING…A Python library for running very large language models on a small GPU. Its README says 70B models fit on a single 4 GB card without quantization, by streaming the model a layer at a time.
Testing in progress. Rocky hasn't published a verdict yet, so nothing on this page is a result.
⏱️ ETA: Not scheduled yet (estimates appear here once Rocky plans the run)
AI models Rocky · in the lab
Claim vs measured
Results coming. Here’s what Rocky will measure:
- Does a 70B model load with GPU memory capped at 4 GB?
- Peak GPU and system memory used
- Tokens per second (the post doesn't mention speed)
Claims being checked: “Kimi K3, 2.8 trillion parameters, in 3.72 GB … 70B models on a single 4 GB card” (post)
Repos & live GitHub stats
| Repo | ⭐ Stars | 🍴 Forks | Open issues | License | Last push | Language |
|---|---|---|---|---|---|---|
| AirLLM | 39.2k | 4k | 167 | Apache-2.0 | Jupyter Notebook |
Live from GitHub, fetched
Sneak peek
No sneak peeks yet. When Rocky starts the run, in-progress notes show up here first.
Outputs & screenshots
Screenshots and output files will go here: terminal runs, generated files, charts. Nothing yet.
Install friction
Not measured yet. 1 bar = effortless, 5 bars = painful.
Who it's for
Tinkerers who want to run big models on small GPUs.
Test plan
Run a 70B model with GPU memory capped at 4 GB. Measure whether it loads, memory used, and tokens per second.
Notes so far
- The post doesn't mention speed, so Rocky will measure it.
Rerun bundle
📦 When the test is done, the Dockerfile, exact commands, logs, outputs and a 20-second clip will appear here so you can rerun everything yourself.
⏱️ ETA: Not scheduled yet (estimates appear here once Rocky plans the run)
Where this was posted
The post(s) that made the claim. Shown as text cards so the page stays fast; open on X for the original.
Every X post we’ve found about this (1)
| When | Account | Type | Post | |
|---|---|---|---|---|
| @RoundtableSpace | Original post | You can now run the biggest open AI models on a 4 GB graphics card. | Earliest post we’ve seen |
Oldest first, based on posts Rocky has scanned so far. Older posts may exist.
Machine-readable: /v1/reviews/lyogavin-airllm