Best Open-Source LLMs 2026: 7 Options Compared

Open source and open weights are not the same thing, and the label on an LLM matters more than it seems. The models here release their weights under an OSI-style licence such as Apache-2.0 or MIT, so you can download them, run them on your own hardware and use them commercially without asking. That excludes some popular families that ship under custom licences with usage limits, which is why you will not see them in this list. Even within the list, read carefully. DeepSeek-R1 itself is MIT, but its smaller distilled versions inherit the licence of the Qwen or Llama model they were built on. OLMo goes further than the others by publishing training data names, code and logs, while most of the rest give you weights only. Size decides cost. A model that needs one 80GB GPU is a serious purchase, a 20B model that fits in 16GB runs on a good laptop, and quantized builds shrink both. The work is picking a runtime, because the weights alone do nothing, and tools like Ollama, llama.cpp and vLLM handle that part. Test two or three on your own prompts before you decide, since benchmark rankings rarely match your task. Free to download is not free to run, so price the hardware.

Qwen3

The widest range of sizes here, from 0.6B to a 235B mixture-of-experts model, all under Apache-2.0. The 8B card shows thinking mode on by default and 32,768 native context tokens, so check how you switch it off.

Price: Free (weights are free to download; you pay for your own hardware or hosting) (Qwen3 pricing, checked 2026-10-23).

Platforms and license: Self-hosted (model weights). Apache-2.0 (open-weight models) (open source) (license source).

Project activity: 27.7k GitHub stars · last commit Jan 9, 2026

Best for: General use across laptop to server sizes.

Worth knowing: Sizes from 0.6B to 235B-A22B (MoE). The README lists Transformers, llama.cpp, Ollama, LM Studio, MLX, vLLM and SGLang as ways to run it. The Qwen3-8B card gives 32,768 native context tokens.

gpt-oss

OpenAI's Apache-2.0 weights: a 120B model that fits one 80GB GPU and a 20B model that runs within 16GB. It must be used with the harmony response format, which the Transformers chat template applies for you.

Price: Free (weights are free to download; you pay for your own hardware or hosting) (gpt-oss pricing, checked 2026-10-23).

Platforms and license: Self-hosted (model weights). Apache-2.0 (open source) (license source).

Project activity: 20.5k GitHub stars · latest release v0.0.9 (Jan 13, 2026) · last commit Oct 12, 2026

Best for: Local agents and tool calling on one GPU.

Worth knowing: gpt-oss-120b (117B total, 5.1B active) fits on a single 80GB GPU with MXFP4; gpt-oss-20b (21B total, 3.6B active) runs within 16GB of memory. Responses use the harmony format.

DeepSeek-R1

MIT-licensed reasoning model at 671B parameters with 37B active, so the full model needs a cluster. The smaller distilled versions are practical, but they carry the Qwen or Llama licence of their base.

Price: Free (weights are free to download; you pay for your own hardware or hosting) (DeepSeek-R1 pricing, checked 2026-10-23).

Platforms and license: Self-hosted (model weights). MIT (distilled models keep their base model's license) (open source) (license source).

Project activity: 92k GitHub stars · latest release v1.0.0 (Jun 27, 2025) · No commits since Jun 27, 2025

Best for: Reasoning-heavy tasks where you can use distilled sizes.

Worth knowing: The full model is 671B parameters (37B active); the 1.5B to 70B distilled versions are built on Qwen2.5 (Apache 2.0) and Llama 3.1/3.3 (Llama licenses), so check the base license before using a distilled one.

Mistral Small

A 24B Apache-2.0 model that, once quantized, fits a single RTX 4090 or a 32GB MacBook per its card. It lists 11 languages, a stronger multilingual choice than the English-first Phi-4.

Price: Free (weights are free to download; you pay for your own hardware or hosting) (Mistral Small pricing, checked 2026-10-23).

Platforms and license: Self-hosted (model weights). Apache-2.0 (open source) (license source).

Best for: Multilingual assistants on one workstation.

Worth knowing: The 24B card says it fits a single RTX 4090 or a 32GB-RAM MacBook once quantized, and needs about 55 GB of GPU memory unquantized. Mistral's hosted API and Le Chat are separate paid products.

Phi-4

A 14B MIT-licensed model with 16K context, trained mostly on English and not intended for multilingual use. Small enough for modest hardware and aimed at reasoning, but its short context limits long documents.

Price: Free (weights are free to download; you pay for your own hardware or hosting) (Phi-4 pricing, checked 2026-10-23).

Platforms and license: Self-hosted (model weights). MIT (open source) (license source).

Best for: English reasoning on modest hardware.

Worth knowing: 14B dense model with 16K context, trained mainly on English; the card says it is not intended for multilingual use.

IBM Granite

Granite 4.0 is Apache-2.0 and comes in micro, tiny, small and 8B sizes, each in base and instruct form. It names tool calling, retrieval and JSON output, but the README gives no memory figures to plan around.

Price: Free (weights are free to download; you pay for your own hardware or hosting) (IBM Granite pricing, checked 2026-10-23).

Platforms and license: Self-hosted (model weights). Apache-2.0 (open source) (license source).

Project activity: 211 GitHub stars · last commit Mar 30, 2026

Best for: Enterprise-style RAG and tool use at small sizes.

OLMo

Apache-2.0 and the most open: code, weights, data names, logs and configs are all published, from 1B to 32B in OLMo 2. The repository says it is out of date and points to OLMo-core for new work.

Price: Free (weights are free to download; you pay for your own hardware or hosting) (OLMo pricing, checked 2026-10-23).

Platforms and license: Self-hosted (model weights). Apache-2.0 (open source) (license source).

Project activity: 6.7k GitHub stars · latest release v0.6.0 (Dec 19, 2024) · last commit Nov 24, 2025

Best for: Research and anyone who needs to see the training recipe.

Worth knowing: The repository lists code, weights, training data names, logs and configs as released. It also says the repo is out of date and points to OLMo-core for the latest work; OLMo 2 sizes listed are 1B, 7B, 13B and 32B.

Which one to pick

• You want one family from laptop size to server size: Qwen3
• You have a single 80GB GPU and want agents: gpt-oss
• You need multilingual support on one workstation: Mistral Small
• You want to study or reproduce training: OLMo
• You want a small English reasoning model: Phi-4