Models small enough to run on your own laptop, scored on the same benchmarks.
Run it on your own laptop
These models are small enough to download and run yourself with Ollama or vLLM - free, offline, and private. The scores are the same Artificial Analysis benchmarks used everywhere else on this site, so they compare directly with the frontier models.
Measured on your machine: ThinkPad L390 Yoga · Intel Core i5-8365U (4 cores, no discrete GPU) · 64GB RAM
Your memory is not the limit - 64GB is far more than any of these need. The limit is the processor. With no separate graphics card, every word is produced on four CPU cores, and speed falls as the model grows. That is why a 30B model loads fine and is still unusable, while a 4B model is genuinely pleasant.
INSTALL THIS ONE
NVIDIA Nemotron 3 Nano 4B
4Bparameters
10/swords per second
1.8intelligence
-coding
It scores 1.8 on intelligence and - on coding, and runs at about 10 words per second on your machine - fast enough to think with rather than wait for.
Artificial Analysis has scored only 1 model this size on intelligence, so there is nothing meaningful to chart. Those models: NVIDIA Nemotron 3 Nano 4B (1.8). That scarcity is itself the answer - laptop-sized models are not where this capability lives yet.
Artificial Analysis has scored only 0 models this size on coding, so there is nothing meaningful to chart. That scarcity is itself the answer - laptop-sized models are not where this capability lives yet.
Artificial Analysis has scored only 0 models this size on agentic, so there is nothing meaningful to chart. That scarcity is itself the answer - laptop-sized models are not where this capability lives yet.
Show up to
Model
Size
Speed on your laptopbandwidth ÷ memory
Memorysize × 0.6 GB
Intelligence
Coding
Agentic
Qwen3.8 27B (Xhigh)
27B
2/s Too slow to enjoy
17.7 GB
20.1
-
-
Qwen3.8 27B (Medium)
27B
2/s Too slow to enjoy
17.7 GB
17.5
-
-
Qwen3.8 27B (Low)
27B
2/s Too slow to enjoy
17.7 GB
15.7
-
-
NVIDIA Nemotron 3 Nano 4B
4B
10/s Comfortable
3.9 GB
1.8
-
-
About the Agentic column: it is nearly empty, and that is the honest picture rather than missing data. Artificial Analysis has scored only three models this size on agentic ability, and all three came out near zero. Models small enough to run on a laptop are genuinely poor at driving tools through a multi-step job - which is exactly the work Claude Code does for you. Use these for writing and answering, not for agent workflows. An earlier version of this site showed flattering agentic numbers here; they were our own arithmetic, ran about 31 points too high against the real Artificial Analysis measure, and have been removed.
Sizes are read from the model name, which is how open-weight models are almost always labelled - Artificial Analysis does not publish parameter counts on its free tier. Speeds are estimated from model size and your memory bandwidth, not measured on your machine, so treat them as a guide. Expect a pause before the first word on long prompts; that part is slower than the numbers here suggest.