Open source models

Open source (and open-weight) models publish downloadable parameters so you can self-host, inspect configuration, fine-tune and deploy inside private infrastructure — subject to the licence.

Executive summary

Open-weight guides in this section focus on what the weights enable, what they do not guarantee (training data, easy reproduction, cheap inference), and how to deploy responsibly. Prefer the OSI label only when the release is genuinely reproducible and permissively licensed.

When to use this page

• You need private inference, fine-tuning or on-prem control.
• You are evaluating licence, hardware and serving-engine requirements.
• You want architecture literacy before a vendor or build decision.

Models in this category

Model	Publisher	Highlights	Guide
Kimi K3	Moonshot AI	~2.8T MoE, ~104B active, 1M context, KDA, custom licence	Kimi K3
Qwen 3.5	Alibaba / Qwen	Hybrid Gated DeltaNet + MoE, 397B-A17B, Apache 2.0	Qwen 3.5
Gemma 4	Google DeepMind	Dense/MoE ladder, local multimodal, Apache 2.0	Gemma 4
Llama 4	Meta	Maverick / Scout MoE, early fusion, Community Licence	Llama 4
DeepSeek	DeepSeek	MLA, DeepSeekMoE, GRPO, V3/R1/V4 long-context	DeepSeek
Mistral	Mistral AI	Small 4 unified MoE, Medium 3.5 dense, Apache 2.0	Mistral
GLM-5	Z.ai	Sparse attention, agentic engineering, Apache 2.0	GLM-5
Nemotron 3	NVIDIA	Hybrid Mamba + LatentMoE, open recipes, NVFP4	Nemotron 3
Command A+	Cohere	218B/25B MoE, RAG/agents, sovereign deploy, Apache 2.0	Command A+
gpt-oss	OpenAI	120B/20B reasoning MoE, MXFP4, Apache 2.0	gpt-oss

How to read an open-weight guide

1. What kind of model — open-weight vs open-data vs open-code; licence implications.
2. Configuration — total vs activated parameters, layers, experts, context.
3. Architecture — attention, MoE, vision, residuals and stability tricks.
4. Training — pre-train → long context → SFT → RL → distillation → quantisation.
5. Deployment — hosted API vs true self-host; GPU shapes; serving engines; ops checklist.

What to do next

Start with Kimi K3 or DeepSeek, compare categories in the AI model landscape 2026, or follow the LLM technical reading map.