Best local models for Ollama and LM Studio
Checked on
Start with Gemma 4 E4B in Ollama (ollama run gemma4:e4b). It is Apache 2.0, a 6.6 GB Q4 download with a 128K context window, and it takes image input. Prefer a desktop app? LM Studio lists google/gemma-4-e4b (also -e2b, -12b, -26b-a4b, -31b, plus -qat). With 24 to 32 GB, step up to a 26B or 27B tag from the table below. Every size and command is from the model cards we checked on Oct 29, 2026.
Also see how to run AI locallyand the API price comparison.
Local model list
21 active open-weight models from our directory. Sizes are the vendor downloads, not our measurements.
| Model | Context | Size | Ollama | LM Studio | License | Good for |
|---|---|---|---|---|---|---|
| Bonsai 2 (local) PrismML | 256K | 2B 911 MB; 4B 1.6 GB; 9B 3.2 GB; 27B 8.3 GB | bonsai-2 (9b default; also 2b, 4b, 27b) | Not in the LM Studio catalog yet (it lists the older Bonsai 27B) | Apache 2.0 | The smallest useful local models with vision and tools. The 9B is a 3.2 GB download for 8 GB laptops; the 2B and 4B suit old machines. |
| Carbon-X 14B (local) Hugging Face | 40,960 tokens | GGUF Q8_0 15.8 GB (recommended); Q4_K_M 9.6 GB; full BF16 weights need one 80 GB GPU with vLLM | Not in the Ollama library | Not in the LM Studio catalog | Apache 2.0 | Asking questions about DNA in plain English on your own hardware: whether a variant looks benign or pathogenic, what a region does, which organism group a sequence comes from. |
| d3 decision models (local) vLLM Semantic Router | 256K | 2B 3.1 GB; 4B 5.6 GB; 9B 10 GB; 27B 29 GB | d3 (4B default; also d3:2b, d3:9b, d3:27b) | Not in the LM Studio catalog yet | Apache 2.0 | Routing support tickets, triaging documents and yes/no or score checks on text, structured data or images, with probabilities instead of a written answer. |
| DeepSeek-R1 distills (local) DeepSeek | 128K | 8B 5.2 GB Q4 / 8.9 GB Q8; 14B 9.0 / 16 GB; 32B 20 GB; 70B 43 GB | deepseek-r1:8b (also 1.5b, 7b, 14b, 32b, 70b) | deepseek/deepseek-r1-0528-qwen3-8b | MIT | Step-by-step reasoning and math on modest hardware. The 8B (R1-0528 on Qwen3) fits a 16 GB laptop. |
| Devstral Small 2 24B (local) Mistral | 384K | 15 GB Q4 / 26 GB Q8 | devstral-small-2:24b | mistralai/devstral-small-2-2512 | Apache 2.0 | Local coding agent: explores codebases, edits multiple files and drives tools. |
| Gemma 4 (local) Google | 128K (E2B, E4B) / 256K (12B, 26B, 31B) | E2B 4.6 GB Q4 / 6.0 GB Q8; E4B 6.6 / 9.0 GB; 12B 8.0 / 13 GB; 26B A4B 18 / 28 GB; 31B 20 / 34 GB | gemma4:e4b (also e2b, 12b, 26b, 31b) | google/gemma-4-e4b (also -e2b, -12b, -26b-a4b, -31b, plus -qat) | Apache 2.0 | Best all-round local family: E4B for laptops, 26B A4B for strong reasoning and code on a 24 GB GPU or 32 GB Mac. |
| gpt-oss (local) OpenAI | 128K | 20B 14 GB; 120B 65 GB (as shipped by Ollama) | gpt-oss:20b (also 120b) | openai/gpt-oss-20b (also openai/gpt-oss-120b) | Apache 2.0 | OpenAI's open-weight reasoning models. 20B runs on a 16 GB GPU or Mac; 120B needs about 80 GB. |
| LFM2.5 24B-A2B (local) Liquid AI | 128K | Q4_0 14 GB; Q4_K_M 15 GB; Q8_0 25 GB; BF16 48 GB | lfm2.5:24b (Q4_K_M; also 24b-a2b-q4_0, -q8_0, -bf16) | Not in the LM Studio catalog yet (it lists the older LFM2-24B-A2B) | LFM Open License v1.0 (free commercial use under $10M annual revenue) | A fast local model for agents on a 24 to 32 GB machine. It has 24B parameters but uses only about 2B per token, so it reasons and calls tools quickly. |
| LFM3-1.7B (local) Liquid AI | 256K tokens | Ollama 1.7 GB (nvfp4); GGUF Q4_K_M 1.1 GB; Q8_0 1.85 GB; BF16 3.47 GB | lfm3 (1.7b; also 1.7b-mxfp8 and 1.7b-mlx-bf16). Mac only in v0.40.12; Linux and Windows since v0.40.13 (Oct 28, 2026) | Not confirmed in the LM Studio catalog yet (Liquid names LM Studio as a launch partner) | LFM Open License v1.0 (commercial use only under $10M annual revenue) | A tiny local model for agents that call tools, fill in structured output or run on a phone or old laptop. About 1 GB at 4-bit. Not for agentic coding or long multi-step reasoning. |
| Llama 3.3 70B (local) Meta | 128K | 43 GB Q4 / 75 GB Q8 | llama3.3:70b | Search "Llama 3.3 70B" in LM Studio (community GGUF) | Llama 3.3 Community License | Proven general model if you have 48 GB or more of memory. Older than Muse Glimmer but widely supported. |
| Ministral 3 (local) Mistral | 256K | 3B 3.0 GB Q4; 8B 6.0 / 9.9 GB; 14B 9.1 / 15 GB | ministral-3:8b (also 3b, 14b) | mistralai/ministral-3-14b | Apache 2.0 | Small edge models (3B, 8B, 14B) with vision for low-memory laptops. |
| Mistral Small 3.2 24B (local) Mistral | 128K | 15 GB Q4 / 26 GB Q8 | mistral-small3.2:24b | mistralai/mistral-small-3.2 | Apache 2.0 | Solid general assistant with vision and reliable function calling on a 24 GB machine. |
| Muse Glimmer 30B (local) Meta | 128K | 18 GB Q4 / 31 GB Q8 / 57 GB BF16 | muse-glimmer:30b | meta/muse-glimmer | Apache 2.0 | Meta's open model for always-on local agents: tool use, long tasks and failure recovery on a single GPU. |
| Nanbeige4.3-33B (local) Nanbeige LLM Lab | 512K tokens (524,288) | Q4_K_M 20.8 GB; Q5_K_M 25.1 GB; Q6_K 29.0 GB; Q8_0 35.2 GB; IQ3_M 17.8 GB | Not in the Ollama library | Search Nanbeige4.3-33B GGUF (bartowski build) | Apache 2.0 | A strong open coding and office-task agent model for a 24 GB GPU or a Mac with 32 GB, if you are comfortable running llama.cpp or LM Studio. |
| Nemotron 3.5 Bolt 74B-A9B (local) NVIDIA | 256K | 25 GB download (Ollama); LM Studio lists 22 GB minimum memory | nemotron-3.5-bolt (74b-a9b) | nvidia/nemotron-3.5-bolt-74b-a9b | OpenMDW 1.1 | NVIDIA's local agent model for a single 24 GB NVIDIA GPU: coding, tool use, and reading screenshots, charts and documents. |
| Nemotron 3.5 Bolt E4B (local) NVIDIA | 128K (64K for the MLX tags) | e4b 7.2 to 7.9 GB; Q8_0 9.8 GB; BF16 17 GB | nemotron-3.5-bolt:e4b (also e4b-q4_K_M, e4b-q8_0, e4b-bf16 and MLX tags) | nvidia/nemotron-3.5-bolt-e4b (5 GB minimum memory) | OpenMDW 1.1 | Fast local voice and device assistants, game characters and small tool-using agents on a 16 GB laptop, Mac or PC. |
| Nemotron 3.5 Super VL 120B (local) NVIDIA | Up to 1M | 80 to 90 GB download (Ollama); LM Studio lists 70 GB minimum memory | nemotron-3.5-super (120b) | nvidia/nemotron-3.5-super-vl | OpenMDW 1.1 | Workstation-class open model for document-heavy agents: OCR, tables, charts and PDFs plus tool calling. |
| NVIDIA Dory 1.0 4B (local) NVIDIA | Up to 256K | Q5_K_M GGUF 3.7 GB; BF16, MXFP8 and NVFP4 on Hugging Face | Not in the Ollama library yet | Not in LM Studio yet | OpenMDW 1.1 | A small reasoning model where you choose how hard it thinks: one to four passes through its shared layers from the same file. For tinkerers comfortable building llama.cpp or vLLM from source. |
| Qwen 3.5 (local) Alibaba Qwen | 256K | 4B 3.3 GB Q4 / 5.2 GB Q8; 9B 6.6 / 10 GB; 27B 17 / 30 GB; 35B 22 GB; 122B 81 GB | qwen3.5:9b (also 0.8b, 2b, 4b, 27b, 35b, 122b) | qwen/qwen3.5-9b (also -2b, -4b, -27b, -35b-a3b) | Apache 2.0 | Wide size range from 0.8B to 122B. The 9B is a strong laptop coder; 27B and 35B for workstation use. |
| Qwen 3.8 27B (local) Alibaba Qwen | 256K | 18 GB Q4 / 30 GB Q8 / 56 GB BF16 | qwen3.8:27b | qwen/qwen3.8-27b | Apache 2.0 | Newest Qwen open model for coding, research and long agentic tasks on a 24 to 32 GB machine. |
| Recursive Flint (local) Recursive | 262K | 23 GB | flint (also flint:40b) | Not in the LM Studio catalog yet | Apache 2.0 | A local coding agent model for a 48 GB Mac: reading unfamiliar repos, tracing bugs, adding features and tests. |
Email me when these prices change
Get notified for the next 14 days whenever an AI lab changes its token pricing or discounts expire.
Free 14-day price change alerts. Zero spam, unsubscribe anytime in one click.
FAQ
What local AI models are available?
Our directory currently lists 21 active local models you can run with Ollama, LM Studio or llama.cpp. The table on this page is the list, with each model's download size, license and run command.
What are the best open-weight models?
For most people, Gemma 4 is the all-round local family: E4B on a 16 GB laptop and 26B A4B on a 24 to 32 GB machine. Qwen 3.5 and Qwen 3.8 are the main alternatives, especially for coding. That is the same starting point as our how to run AI locally setup. Cloud models still lead on the hardest tasks.
Where can I find an AI model list?
This page is the list of local, off-grid models in our directory. For hosted APIs, see the AI model API price comparison. Each row links to a model page with the official source we checked.
What off-grid AI models can I run?
Any row in the table. They run on your machine with no API bill. Plan on RAM or VRAM of roughly the download size plus 2 to 4 GB for context. Some tags are NVIDIA-only or not in Ollama or LM Studio yet; the row says so.
Does Ternary Bonsai 2 27B work in LM Studio?
Bonsai 2 is in the Ollama library (bonsai-2 (9b default; also 2b, 4b, 27b)). Not in the LM Studio catalog yet (it lists the older Bonsai 27B). The 27B tag is an 8.3 GB download. Apache 2.0.
Sources
- Bonsai 2 (local) source
- Carbon-X 14B (local) source
- d3 decision models (local) source
- DeepSeek-R1 distills (local) source
- Devstral Small 2 24B (local) source
- Gemma 4 (local) source
- gpt-oss (local) source
- LFM2.5 24B-A2B (local) source
- LFM3-1.7B (local) source
- Llama 3.3 70B (local) source
- Ministral 3 (local) source
- Mistral Small 3.2 24B (local) source
- Muse Glimmer 30B (local) source
- Nanbeige4.3-33B (local) source
- Nemotron 3.5 Bolt E4B (local) source
- Nemotron 3.5 Super VL 120B (local) source
- NVIDIA Dory 1.0 4B (local) source
- Qwen 3.5 (local) source
- Qwen 3.8 27B (local) source
- Recursive Flint (local) source
We do not invent sizes or prices. If a vendor page disagrees with a row, trust the vendor and tell us.