Small on-device LLMs

Black Forest Labs · 4B · Apache 2.0

Fastest open FLUX.2 — sub-second text-to-image and multi-reference editing on consumer GPUs

Alibaba · 5B · Apache 2.0

Unified text/image-to-video — the local sweet spot under Apache 2.0

Alibaba · 6B · Apache 2.0

8-step distilled image model — photorealism and bilingual text on 16GB cards

Tencent · 8.3B · Tencent Hunyuan Community

Compact cinematic video model — strong faces and motion on a single 4090

Alibaba · 8.8B · Apache 2.0

The community-favourite local VLM — superb OCR, receipts & captioning

Alibaba · 9B · Apache 2.0

Multimodal Qwen 3.5 mid-size

Lightricks · 19B · LTX-2 Community

Open 4K video with native stereo audio — text, image and video-to-video

Alibaba · 20B · Apache 2.0

Open text-to-image with strong English and Chinese typography

OpenAI · 21B · Apache 2.0

OpenAI's open-weight MoE with configurable reasoning

Google · 27B · Gemma

Gemma 4 MoE instruct model (official)

Alibaba · 27B · Apache 2.0

Flagship dense Qwen 3.8 — native multimodal all-rounder with video understanding

Alibaba · 27B · Apache 2.0

Flagship open Wan 2.2 — 14B-active MoE for photoreal text-to-video

Meta · 30B · Apache 2.0

Open agentic 30B distilled from Muse Spark — tool use, vision and local recovery on a single GPU

Alibaba · 31B · Apache 2.0

Efficient vision MoE — 3B active, strong temporal & document understanding

Black Forest Labs · 32B · FLUX Non-Commercial

Flagship open-weight FLUX.2 — text-to-image and multi-reference editing up to 4MP

MiniMax · 33B · MiniMax Community

Open video generation — text/image to 2K video with native stereo audio

InternScience · 35B · Apache 2.0

Efficient multimodal agentic MoE for long-horizon search, engineering and scientific research

DeepReinforce · 35B · MIT

Agentic coding MoE with a 3B active working set and self-improving training

Alibaba · 36B · Apache 2.0

Big-model quality at 3B-active speed — the mid-hardware sweet spot

Tencent · 80B · Tencent Hunyuan Community

Reasoning image model — prompt rewrite, chain-of-thought and image-to-image editing

Meta · 109B · Llama 4 Community

MoE with 16 experts, 17B active params

OpenAI · 117B · Apache 2.0

OpenAI's flagship open-weight MoE — 52.6% SWE-bench

Mistral AI · 119B · Apache 2.0

Sparse Mistral Small 4 — 6.5B active, strong local all-rounder

Alibaba · 235B · Apache 2.0

Flagship vision-language MoE — frontier multimodal reasoning and agentic GUI control

Tencent · 295B · Apache 2.0

Production-focused agentic MoE with strong coding, tool use and long-context reasoning

MiniMax · 428B · MiniMax Community

Native multimodal MoE — understands text, image and long video with 1M context

Z.ai · 753B · MIT

Same 753B / 40B-active base as GLM-5.2 — post-training lifts coding and long-horizon agents, 1M context

Meituan · 1.6T · MIT

Frontier-scale agentic and coding MoE with sparse attention and native 1M context

DeepSeek · 1.6T · MIT

Flagship V4 MoE — 49B active, 1M context

Alibaba · 2.4T · Qwen

Frontier Qwen 3.8 MoE — 95B active, 1M context

Moonshot AI · 2.8T · Kimi

Frontier 2.8T multimodal MoE — 104B active, native video understanding, 1M context

Alibaba · 0.6B · Apache 2.0

Ultra-light Qwen 3 model for constrained devices

Alibaba · 0.8B · Apache 2.0

Ultra-tiny model for embedded and edge

Meta · 1B · Llama 3.2 Community

Meta's smallest Llama for edge devices

Google · 1B · Gemma

Google's tiny Gemma for on-device

Alibaba · 1.3B · Apache 2.0

Tiny open text-to-video — 480p clips on 8GB consumer GPUs

Alibaba · 1.5B · Apache 2.0

Ultra-lightweight coding model

DeepSeek · 1.5B · MIT

Tiny reasoning model distilled from R1

Alibaba · 1.7B · Apache 2.0

Compact multilingual Qwen 3

Alibaba · 2B · Apache 2.0

Small multimodal Qwen 3.5

Meta · 3B · Llama 3.2 Community

Lightweight Llama for mobile and edge

HuggingFace · 3B · Apache 2.0

Lightweight multilingual reasoning

IBM · 3B · Apache 2.0

Compact enterprise model for edge and constrained environments

Mistral AI · 3B · Apache 2.0

Current-gen tiny Ministral — edge chat with 256K context

Microsoft · 3.8B · MIT

Lightweight reasoning model

Google · 4B · Gemma

Multimodal Gemma with 128K context

Alibaba · 4B · Apache 2.0

Small multimodal Qwen 3.5

Alibaba · 4.4B · Apache 2.0

Compact dedicated vision-language model — OCR & image chat on edge

Google · 5B · Gemma

Gemma 4 efficient instruct model (official)

Alibaba · 7B · Apache 2.0

Dedicated coding model

DeepSeek · 7B · MIT

R1 reasoning distilled into Qwen 7B

Google · 8B · Gemma

Gemma 4 balanced instruct model (official)

Meta · 8B · Llama 3.1 Community

Meta's versatile 8B — great quality/speed ratio

Alibaba · 8B · Apache 2.0

Qwen 3 with thinking mode support

IBM · 8B · Apache 2.0

Balanced general-purpose enterprise model

Mistral AI · 8B · MRL

Mistral's efficient 8B model

Zhipu AI · 9B · GLM-4

Multilingual model supporting 26 languages with 128K context

NVIDIA · 9B · NVIDIA Open

Hybrid Mamba2 architecture for reasoning

DeepReinforce · 9B · MIT

Self-improving agentic coding model optimized for terminal and software engineering tasks

Black Forest Labs · 9B · FLUX Non-Commercial

Higher-quality distilled FLUX.2 — sub-second generation and multi-reference editing

Google · 12B · Gemma

Multimodal Gemma with 128K context

Mistral AI · 12B · Apache 2.0

Multilingual 12B with 128K context

Google · 12B · Apache 2.0

Gemma 4 mid-size instruct — multimodal any-to-any

Microsoft · 14B · MIT

Microsoft's reasoning-focused model

Alibaba · 14B · Apache 2.0

Strong all-rounder with thinking mode

DeepSeek · 14B · MIT

R1 reasoning distilled into Qwen 14B

Mistral AI · 14B · Apache 2.0

Current-gen Ministral mid-size — local assistant with 256K context

Liquid AI · 24B · Liquid AI

Hybrid MoE with convolution+attention layers — 2.3B active

Mistral AI · 24B · Apache 2.0

Coding-focused model with 256K context — 68% SWE-bench

Mistral AI · 24B · Apache 2.0

Multimodal Mistral with vision support

Google · 26B · Apache 2.0

Discrete diffusion MoE — 1100+ tok/s on H100, multimodal (text/image/video)

Alibaba · 27.8B · Apache 2.0

Flagship native multimodal Qwen 3.5

Alibaba · 27.8B · Apache 2.0

Flagship dense Qwen 3.6 — native multimodal all-rounder

Alibaba · 30B · Apache 2.0

MoE with only 3.3B active — extremely efficient

NVIDIA · 30B · NVIDIA Open

MoE with 1M context and 3B active

IBM · 30B · Apache 2.0

High-capacity enterprise model for complex reasoning and tool use

Cohere · 30B · Apache 2.0

Open agentic coding MoE with 3B active — built for software engineering and terminal tasks

Alibaba · 30B · Apache 2.0

Efficient agentic coding MoE — 3B active, 256K context

Alibaba · 32B · Apache 2.0

Qwen 3 flagship dense model

DeepSeek · 32B · MIT

R1 reasoning distilled into Qwen 32B — sweet spot

Allen AI · 32B · Apache 2.0

Fully open research model by Allen AI

Google · 33B · Gemma

Gemma 4 flagship instruct model (official)

Google · 33B · Gemma

Gemma 4 flagship base model (official)

Cohere · 35B · CC BY-NC 4.0

Optimized for retrieval-augmented generation

Alibaba · 35B · Apache 2.0

Efficient multimodal MoE with 3B active

Mistral AI · 47B · Apache 2.0

MoE with 12.9B active params

Meta · 70B · Llama 3.3 Community

Best open model at 70B class

Alibaba · 80B · Apache 2.0

High-sparsity MoE — extreme low activation ratio for fast inference at 80B scale

Alibaba · 80B · Apache 2.0

Ultra-efficient agentic coding MoE optimized for tool-calling coding agents

Tencent · 80B · Tencent Hunyuan Community

Largest open image MoE — 13B active, strong long-prompt generation

Z.ai · 106B · MIT

Consumer-friendly GLM MoE — 12B active, strong agentic & tool use

Alibaba · 122B · Apache 2.0

Large multimodal MoE

DeepSeek · 158B · MIT

Efficient long-context V4 — 13B active, 1M context

Alibaba · 235B · Apache 2.0

Massive MoE with 22B active — frontier quality

Z.ai · 357B · MIT

Large GLM MoE with strong coding and 200K context

Alibaba · 397B · Apache 2.0

Largest multimodal Qwen 3.5 MoE

Meta · 400B · Llama 4 Community

Multimodal MoE with 128 experts — 17B active, 1M context

Alibaba · 480B · Apache 2.0

Largest open coding MoE — 35B active

DeepSeek · 671B · MIT

Massive MoE reasoning model — 37B active

DeepSeek · 685B · MIT

State-of-the-art MoE — 37B active params