Best small open models you can run on a laptop
You do not need a server to run a good open model. These models fit in 16 to 32 GB of memory, or less, and cover chat, reasoning, coding and vision. The memory figures are our estimates at 4-bit precision for the smallest variant, except for models shipped in a lower-bit format, where we quote the file size.
-
Gemma 4
Google's Apache 2.0 family has sizes for almost any machine: phone-sized E2B and E4B, a 12B, a 26B mixture of experts that activates 3.8B, and a 31B flagship. The smallest needs about 3.5 GB at 4-bit by our estimate.
-
gpt-oss
The 21B version of OpenAI's Apache 2.0 reasoning model fits a 16 GB machine because only 3.6B parameters are active per token. It is from 2025, but nearly every local tool supports it.
-
Qwen3.5
Alibaba's Apache 2.0 family has small sizes at 0.8B, 2B, 4B and 9B next to larger ones, all with image and video input and a 256K context. By our estimate the 0.8B needs about 0.6 GB and the 9B about 7 GB at 4-bit.
-
Qwen3.6
A 35B mixture of experts that activates only 3B per token, and a 27B dense model, both under Apache 2.0 and tuned for agentic coding. By our estimate they need about 25 GB and 19 GB at 4-bit, so they suit a 32 GB machine.
-
Laguna XS 2.1
Poolside's 33B coding model that activates only 3B parameters per token. The official Q4_K_M file is 20.3 GB and the card says a Mac with 36 GB can run it, but it warned about llama.cpp, vLLM and Ollama rough edges, so check the current state first.
-
GLM-4.7-Flash
Z.ai's MIT-licensed 30B mixture of experts that activates 3B parameters per token. Official weights are 62.5 GB in BF16 only, so run a community 4-bit build, which needs about 22 GB by our estimate, on a 32 GB Mac or a 24 GB GPU.
-
Raon-1 mini preview
KRAFTON's 28.8B mixture of experts activates only 4.3B parameters per token and was trained from scratch for Korean chat and agents. By our estimate it needs about 18 GB at 4-bit, and its 17.6 GB NVFP4 build fits one 32 GB GPU. Apache 2.0, and the card calls it a preview.
-
Nemotron 3 Nano 4B
NVIDIA's 4B Mamba-Transformer hybrid, compressed from Nano 9B v2, with switchable reasoning for RTX PCs and Jetson boards. By our estimate it needs under 3 GB at 4-bit.
-
GigaChat 3.1 Lightning
Sber's MIT-licensed 10B mixture of experts with 1.8B active parameters, strong in Russian. The FP8 files are 12.3 GB and a Q4_K_M GGUF is 6.5 GB, but all 10B parameters must still fit in memory.
-
Dory 1.0
NVIDIA's 4B looped model: you choose whether its shared block runs one to four times, trading speed for quality from one checkpoint. It is about 8 GB in BF16 and 3.2 GB in NVFP4, with a 3.7 GB GGUF, and OpenMDW-1.1 allows commercial use.
-
Nanbeige4.2-3B
BOSS Zhipin's 3B-class agent model (4.17B parameters with embeddings) under Apache 2.0, with a 256K context. Its own table reports 63.6 on SWE-bench Verified against 53.1 for Qwen3.5-9B, but it needs Nanbeige's forks of vLLM, SGLang or llama.cpp. By our estimate it needs about 3 GB at 4-bit.
-
S5-nano
Maritaca AI's 4B model built on Qwen3.5-4B for Brazilian Portuguese, with Think and Instruct modes, tool calling and a 256K context, under Apache 2.0. The Q4_K_M GGUF is 2.8 GB, and the benchmark scores are Maritaca's own.
-
Spark-X2.6
TokenSpark's 1.7B and 4B models under Apache 2.0, with switchable thinking and a 1M-token position limit, though a context that long needs a lot of memory. The 4B is 3.41 GB in INT4, and its GGUF repository had about 660,000 downloads in the 30 days to 11 November 2026. The benchmark charts are TokenSpark's own.
-
Raon-1 nano
KRAFTON's 3.1B on-device model (1.5B effective), distilled from its larger mixture of experts for Korean and English with tool calling and a 262K context. The 4-bit GGUF is 2.1 GB, and MLX and LiteRT-LM builds are about 1.8 GB. Apache 2.0, and every score is KRAFTON's own.
-
MiniCPM5-2B
A dense 2B model for phones and laptops with a 128K context, tool use and coding, under Apache 2.0, and OpenBMB released its training data. A 1B version is also available.
-
Ternary Bonsai 2
PrismML's Qwen-based models with ternary weights and vision: a 27B-class model in 5.95 GB and a 2B model in 0.43 GB, under Apache 2.0. Some of the file formats need PrismML's llama.cpp fork.
-
LFM2.5-24B-A2B
A 24B mixture of experts that activates only 2.45B parameters per token, so it generates quickly. It needs about 17 GB at 4-bit by our estimate, and the license is free only below US$10 million of revenue.
-
Granite 4.2
IBM's dense reasoning family in 3B, 8B and 30B sizes, with switchable thinking and tool calling, under Apache 2.0 and aimed at enterprise use.
-
Mellum2.1
JetBrains' 12B mixture of experts with 2.5B active parameters, built for coding agents, under Apache 2.0. JetBrains publishes 8.1 GB 4-bit and 7.0 GB MXFP4 files, and its own table shows Qwen3.5-9B ahead on SWE-bench Verified (50.0 against 47.0).
-
Olmo 3.1
Ai2's fully open 7B and 32B models: weights, training data, code and checkpoints are all public, under Apache 2.0. The 7B needs about 5 GB at 4-bit by our estimate.
-
K2 Horizon
IFM's family has a 7B (5.6 GB as 4-bit GGUF), a 3.7B (3.2 GB) and a 0.9B (0.67 GB) under Apache 2.0, with a 512K context (128K for the 0.9B), plus a 36B-A4B that needs about 22 GB in 4-bit. GGUF files need a recent llama.cpp, and in IFM's own table the 7B scores 72.4 on SWE-bench Verified against 50.8 for Qwen3.5-9B.
-
Jais 3
The Arabic-first branch of K2 Horizon, with 0.9B, 3.7B and 7B models for phones and laptops (the 7B is 5.6 GB as 4-bit GGUF) and a 36B-A4B for a workstation. IFM's own tests put the 7B at 79.4 on Arabic against 76.9 for Gemma 4 12B.
-
LFM3
Liquid AI's 2B-class family: a 1.7B text model with a 256K context that fits in about 1 GB, plus vision, audio and omni siblings. The license is free only below US$10 million of revenue.
-
MiniCPM-V 4.7
OpenBMB's vision-language models for images and video: a 1B model that runs on a phone, and a 35B mixture of experts that activates 3B. Apache 2.0.
-
MiniCPM-o 4.5
OpenBMB's 9B model for live video and voice chat, with an 8.5 GB AWQ int4 build, a 12 GB GPU for llama.cpp-omni and Apple silicon support. Speech is English and Chinese.
-
Mage-VL
Microsoft's 4B vision-language model for images, long video and live streams, under Apache 2.0. It reads video like a codec, cutting visual tokens by more than 75%, and Microsoft's tables put it ahead of Qwen3-VL-4B on most video tests. By our estimate it needs about 3 GB at 4-bit.
How we chose
- The smallest variant runs in roughly 16 GB of memory or less, or the model activates few parameters per token.
- The license allows use by most individuals and small companies; the entry notes where it does not.
- Weights are available today on Hugging Face and supported by common local tools.