AIUpdateWatch Daily Intelligence

Local AI models: complete practical comparison

A comprehensive tracked list of major official general-purpose, multimodal and reasoning models up to 32B that can realistically run on consumer or workstation hardware after suitable quantization.

Morning edition: Friday, July 24, 2026 · Data cutoff Jul 23, 2026, 8:30 PM (America/New_York)

Local model

Runs on hardware you control

A useful local model must fit your actual computer at an acceptable speed. Downloadable weights alone do not make a model practical.

Comparison scope

33 official models

This table covers major official chat, instruct, multimodal and reasoning checkpoints up to 32B. It excludes unofficial fine-tunes, base-only checkpoints, embeddings, image generators, audio-only models and data-center models.

Choose by hardware first

Practical local deployment tiers

Compare local hardware →

Very light4–8 GB

Tiny assistants, classification and basic text work.

4 models in table

Light8–12 GB

Everyday local chat, summaries and smaller multimodal tasks.

10 models in table

Mainstream12–16 GB

The strongest practical range for many consumer systems.

8 models in table

Workstation20–24 GB

Stronger reasoning and multimodal work with more capable hardware.

6 models in table

High-end local32–40 GB+

Large quantized models for high-memory PCs and Macs.

5 models in table

Memory bands and comfortable-hardware guidance are planning estimates, not vendor guarantees. They assume a sensible 4-bit or 5-bit quantized build, short-to-moderate context and room for the operating system. Long context, vision input, larger batches and GPU offload can require substantially more memory.

Complete tracked list

Local model comparison table

33 models

Rows are ordered from the lightest deployment tier to the most demanding. Every model name opens the provider’s official model card or documentation. The final column describes hardware for comfortable everyday use rather than the absolute minimum needed to load a model.

Comparison of major official local-capable AI models
Tier	Provider	Model	Parameters	Context	Input	Practical memory	Best for	License	Comfortable hardware
Very light	Google	Gemma 3 270M — open official source in a new tab	0.27B	32K	Text	4 GB+	Classification, extraction and tiny assistants	Gemma terms	Modern 4-core CPU or integrated GPU; 16 GB system or unified memory preferred.
Very light	Qwen	Qwen3-0.6B — open official source in a new tab	0.6B	32K	Text	4 GB+	Light multilingual chat and agent experiments	Apache 2.0	Modern 4-core CPU or integrated GPU; 16 GB system or unified memory preferred.
Very light	Google	Gemma 3 1B — open official source in a new tab	1B	32K	Text	4–8 GB	Light chat, summaries and rewriting	Gemma terms	Modern 4-core CPU or integrated GPU; 16 GB system or unified memory preferred.
Very light	Meta	Llama 3.2 1B Instruct — open official source in a new tab	1B	128K	Text	4–8 GB	Mobile, edge and basic assistant tasks	Llama 3.2	Modern 4-core CPU or integrated GPU; 16 GB system or unified memory preferred.
Light	DeepSeek	DeepSeek-R1-Distill-Qwen-1.5B — open official source in a new tab	1.5B	32K recommended	Text · reasoning	8 GB+	Small local math and reasoning experiments	MIT	Modern 6–8 core CPU or 6–8 GB VRAM; 16 GB system or unified memory.
Light	Qwen	Qwen3-1.7B — open official source in a new tab	1.7B	32K · 131K with YaRN	Text	8 GB+	Multilingual chat and efficient reasoning	Apache 2.0	Modern 6–8 core CPU or 6–8 GB VRAM; 16 GB system or unified memory.
Light	IBM	Granite 3.3 2B Instruct — open official source in a new tab	2B	128K	Text	8 GB+	RAG, enterprise text and structured answers	Apache 2.0	Modern 6–8 core CPU or 6–8 GB VRAM; 16 GB system or unified memory.
Light	Meta	Llama 3.2 3B Instruct — open official source in a new tab	3B	128K	Text	8 GB+	General local chat and document assistance	Llama 3.2	Modern 6–8 core CPU or 6–8 GB VRAM; 16 GB system or unified memory.
Light	Hugging Face	SmolLM3-3B — open official source in a new tab	3B	64K · 128K with YaRN	Text	8 GB+	Compact chat, tools and local application use	Apache 2.0	Modern 6–8 core CPU or 6–8 GB VRAM; 16 GB system or unified memory.
Light	Mistral AI	Ministral 3 3B Instruct 2512 — open official source in a new tab	3.4B + vision encoder	262K maximum	Text + image	8 GB FP8 · less quantized	Multimodal edge and private local assistants	Apache 2.0	Modern 6–8 core CPU or 6–8 GB VRAM; 16 GB system or unified memory. Keep extra memory headroom for multimodal input.
Light	Microsoft	Phi-4-mini-instruct — open official source in a new tab	3.8B	128K	Text	8–12 GB	Reasoning, multilingual use and function calling	MIT	Modern 6–8 core CPU or 6–8 GB VRAM; 16 GB system or unified memory.
Light	Qwen	Qwen3-4B — open official source in a new tab	4B	32K · 131K with YaRN	Text	8–12 GB	Balanced multilingual chat and reasoning	Apache 2.0	Modern 6–8 core CPU or 6–8 GB VRAM; 16 GB system or unified memory.
Light	Google	Gemma 3 4B — open official source in a new tab	4B	128K	Text + image	8–12 GB	Compact multimodal assistance	Gemma terms	Modern 6–8 core CPU or 6–8 GB VRAM; 16 GB system or unified memory. Keep extra memory headroom for multimodal input.
Light	Google	Gemma 3n E2B Instruct — open official source in a new tab	6B total · E2B effective	32K	Text + image + audio + video	8 GB+ with offload	Low-resource multimodal and mobile-device work	Gemma terms	Modern 6–8 core CPU or 6–8 GB VRAM; 16 GB system or unified memory. Keep extra memory headroom for multimodal input.
Mainstream	Google	Gemma 3n E4B Instruct — open official source in a new tab	8B total · E4B effective	32K	Text + image + audio + video	12 GB+ with offload	Broader local multimodal understanding	Gemma terms	Modern 8-core CPU with 8–16 GB VRAM, or 24–32 GB unified memory. Keep extra memory headroom for multimodal input.
Mainstream	DeepSeek	DeepSeek-R1-Distill-Qwen-7B — open official source in a new tab	7B	32K recommended	Text · reasoning	12–16 GB	Local math, coding and step-by-step reasoning	MIT	Modern 8-core CPU with 8–16 GB VRAM, or 24–32 GB unified memory.
Mainstream	DeepSeek	DeepSeek-R1-Distill-Llama-8B — open official source in a new tab	8B	32K recommended	Text · reasoning	12–16 GB	Reasoning with a Llama-based checkpoint	MIT	Modern 8-core CPU with 8–16 GB VRAM, or 24–32 GB unified memory.
Mainstream	Qwen	Qwen3-8B — open official source in a new tab	8.2B	32K · 131K with YaRN	Text	12–16 GB	Strong multilingual chat, tools and reasoning	Apache 2.0	Modern 8-core CPU with 8–16 GB VRAM, or 24–32 GB unified memory.
Mainstream	IBM	Granite 3.3 8B Instruct — open official source in a new tab	8B	128K	Text	12–16 GB	Enterprise RAG, coding and structured reasoning	Apache 2.0	Modern 8-core CPU with 8–16 GB VRAM, or 24–32 GB unified memory.
Mainstream	Meta	Llama 3.1 8B Instruct — open official source in a new tab	8B	128K	Text	12–16 GB	Established general-purpose local assistant use	Llama 3.1	Modern 8-core CPU with 8–16 GB VRAM, or 24–32 GB unified memory.
Mainstream	Mistral AI	Ministral 3 8B Instruct 2512 — open official source in a new tab	8.4B + vision encoder	262K maximum	Text + image	12 GB FP8 · less quantized	Stronger multimodal edge deployment	Apache 2.0	Modern 8-core CPU with 8–16 GB VRAM, or 24–32 GB unified memory. Keep extra memory headroom for multimodal input.
Mainstream	OpenAI	gpt-oss-20b — open official source in a new tab	21B total · 3.6B active	128K	Text · reasoning	16 GB native MXFP4	Local reasoning, agents, tools and structured output	Apache 2.0	Modern 8-core CPU with 8–16 GB VRAM, or 24–32 GB unified memory.
Workstation	Google	Gemma 3 12B — open official source in a new tab	12B	128K	Text + image	20–24 GB	Higher-quality multimodal local work	Gemma terms	16–24 GB VRAM, or 32–48 GB unified memory; fast SSD and strong cooling recommended. Keep extra memory headroom for multimodal input.
Workstation	Mistral AI	Mistral Nemo 12B Instruct — open official source in a new tab	12B	128K	Text	20–24 GB	Multilingual chat, code and function calling	Apache 2.0	16–24 GB VRAM, or 32–48 GB unified memory; fast SSD and strong cooling recommended.
Workstation	Microsoft	Phi-4 — open official source in a new tab	14B	16K	Text	20–24 GB	Math, logic and dense reasoning tasks	MIT	16–24 GB VRAM, or 32–48 GB unified memory; fast SSD and strong cooling recommended.
Workstation	Qwen	Qwen3-14B — open official source in a new tab	14B	32K · 131K with YaRN	Text	20–24 GB	Stronger multilingual reasoning and agents	Apache 2.0	16–24 GB VRAM, or 32–48 GB unified memory; fast SSD and strong cooling recommended.
Workstation	DeepSeek	DeepSeek-R1-Distill-Qwen-14B — open official source in a new tab	14B	32K recommended	Text · reasoning	20–24 GB	Stronger local reasoning, math and coding	MIT	16–24 GB VRAM, or 32–48 GB unified memory; fast SSD and strong cooling recommended.
Workstation	Mistral AI	Ministral 3 14B Instruct 2512 — open official source in a new tab	14B + vision encoder	262K maximum	Text + image	20–24 GB quantized	High-quality multimodal workstation use	Apache 2.0	16–24 GB VRAM, or 32–48 GB unified memory; fast SSD and strong cooling recommended. Keep extra memory headroom for multimodal input.
High-end local	Mistral AI	Mistral Small 3.1 24B Instruct — open official source in a new tab	24B	128K	Text + image	32 GB quantized	High-quality private assistants and long documents	Apache 2.0	24–48 GB VRAM, or 48–64 GB unified memory; high-end workstation and possible multi-GPU use. Keep extra memory headroom for multimodal input.
High-end local	Google	Gemma 3 27B — open official source in a new tab	27B	128K	Text + image	32–40 GB	High-end multimodal work and coding	Gemma terms	24–48 GB VRAM, or 48–64 GB unified memory; high-end workstation and possible multi-GPU use. Keep extra memory headroom for multimodal input.
High-end local	Qwen	Qwen3-30B-A3B — open official source in a new tab	30B total · 3B active	32K · 131K with YaRN	Text	32 GB+	Efficient MoE reasoning, agents and multilingual work	Apache 2.0	24–48 GB VRAM, or 48–64 GB unified memory; high-end workstation and possible multi-GPU use.
High-end local	Qwen	Qwen3-32B — open official source in a new tab	32B	32K · 131K with YaRN	Text	40 GB+	High-end multilingual reasoning and agent tasks	Apache 2.0	24–48 GB VRAM, or 48–64 GB unified memory; high-end workstation and possible multi-GPU use.
High-end local	DeepSeek	DeepSeek-R1-Distill-Qwen-32B — open official source in a new tab	32B	32K recommended	Text · reasoning	40 GB+	The strongest practical DeepSeek R1 distill	MIT	24–48 GB VRAM, or 48–64 GB unified memory; high-end workstation and possible multi-GPU use.

How to choose from the table

Start below your limit

A model that barely fits leaves too little memory for context, applications and the operating system. Choose one tier below your theoretical maximum for a smoother experience.

Do not chase context numbers

Published 128K or 262K context limits do not mean a consumer computer can use the full window comfortably. Begin with 4K–16K and increase only when necessary.

Test the exact quantization

Quality, speed and memory use vary between GGUF, MLX, ONNX, FP8, MXFP4 and other packages. Evaluate the exact file and runtime you plan to keep.

Need very large self-hosted models?

Open models

See open-weight releases that require server-class or data-center hardware.

Open the Open Models page →

Choosing a computer?

Local hardware

Compare memory, published prices, system fit and hardware cautions before buying.

Open the Local Hardware page →