Last reviewed: September 2026. Model families move fast; I re-check this page every quarter and note changes in the log at the end.
Every leaderboard tells you which model scored highest last week. None of them tells you which model you may ship, which one fits your hardware, or which family will still be maintained in a year. This page is the map I actually use when I pick a model for a client: families rather than point releases, the license named for each, and capability described in words that will still be true after the next version.
This is chapter 4 of the Open Source AI Hub. The license terms mentioned here are read in full in chapter 3, and AILane will tell you which of these families your own machine can run.
How to read this page
Four things describe a model well enough to choose it. Size is the number of parameters, and for mixture-of-experts models the number that matters is the active parameters used per token, which is much smaller than the total; a “235B” mixture-of-experts with 22B active runs more like a 22B dense model in speed and more like a much larger one in memory. Context length is how much text the model can consider at once, and the advertised figure is usually far above the length at which quality holds up. License is which of the families in chapter 3 the weights fall into. And the base-versus-instruct distinction: a base model completes text and is a starting point for training, an instruct model has been taught to follow instructions and is what you deploy.
The families, by maker
Meta Llama
The family that made open weights mainstream, from Llama 2 in 2023 through the Llama 3 line, with 3.1 adding a 405B model and long context, 3.2 adding small and vision models, and 3.3 a strong 70B, to Llama 4 in April 2025, which moved to mixture-of-experts and native multimodality. License: Llama Community License, so not open source, with the EU restriction on the multimodal models. Strengths: an enormous ecosystem of fine-tunes, tooling support everywhere, and reliable general-purpose quality. I still deploy Llama 3.3 70B where a client wants a well-understood workhorse.
Mistral
Mistral 7B in 2023 proved that a small model could punch above its weight, and Mixtral popularized mixture-of-experts for open weights. The current open line is Mistral Small, a 24B-class dense model that is my default for a single 24 GB GPU; Devstral, a coding agent model; and Magistral Small for reasoning. License: Apache-2.0 for the open line, but the larger models sit under research or non-production licenses, so check per model. Strengths: efficiency, European provenance, good multilingual quality.
Alibaba Qwen
Qwen has become the most complete open family: Qwen2.5 and then Qwen3, released in April 2025, span everything from 0.6B to a 235B mixture-of-experts, with dense models at the sizes people actually run and a hybrid mode that can reason step by step or answer directly. Qwen3-Coder is the strongest open coding family I have used, and the Qwen3 embedding and reranker models are excellent. License: Apache-2.0 across Qwen3. Strengths: breadth, multilingual coverage including Indian languages, and a release cadence you can plan around. If I could keep only one family for client work today it would be this one.
DeepSeek
DeepSeek V3, a 671B mixture-of-experts with 37B active parameters, and DeepSeek R1, its reasoning counterpart released in January 2025, showed that frontier-class open weights could come from a small team at a fraction of the assumed cost. The distilled R1 models at 7B to 70B are what most people actually run. License: MIT for R1 and for the V3 line from March 2025 onward. Strengths: reasoning and mathematics, cost-efficient serving at scale; the full models need serious hardware.
Google Gemma
Gemma 3, released in March 2025, comes in 1B, 4B, 12B and 27B sizes with vision in the larger ones, and the 27B is competitive with much bigger models. Gemma 3n targets phones and laptops. License: Gemma Terms of Use, so not open source. Strengths: quality per parameter, strong multilingual results, and a 27B that fits a single consumer GPU at 4-bit.
Microsoft Phi
The Phi family, from Phi-3 through Phi-4 (December 2024) and the Phi-4-mini and reasoning variants in 2025, is built on curated and synthetic data to get large-model behavior from small models, at 4B to 14B. License: MIT. Strengths: small-footprint reasoning and math; weaker general knowledge than models trained on the open web.
OpenAI gpt-oss
OpenAI returned to open weights on 5 August 2025 with gpt-oss-120b and gpt-oss-20b, both mixture-of-experts reasoning models with adjustable reasoning effort. The 20B runs on a 16 GB machine and the 120B on a single 80 GB GPU. License: Apache-2.0, with a short usage policy. Strengths: strong reasoning and tool use for their size; no data information, so open-weight rather than open source.
Ai2 OLMo
OLMo is the reference for fully open models: weights, training code, the Dolma corpus and the training logs are all published, so it is one of the few families that meets the Open Source AI Definition outright. Quality trails the leaders at the same size but is entirely respectable. License: Apache-2.0. Strengths: auditability and reproducibility; the right choice when a client needs to be able to say exactly what the model learned from.
IBM Granite
Granite models are enterprise-oriented small and mid-sized models with unusually careful disclosure of training data sources, and IBM indemnifies its own customers for their use. License: Apache-2.0. Strengths: procurement-friendly provenance; a sensible default for regulated industries that want a permissive license and a vendor they can name.
Others worth knowing
NVIDIA’s Nemotron models, tuned for agentic and reasoning use on NVIDIA hardware, under NVIDIA’s own open model license or the Llama license depending on the base. Moonshot’s Kimi K2, a trillion-parameter mixture-of-experts strong at agentic coding, under a modified MIT license. Zhipu’s GLM line, competitive open-weight models under permissive terms in recent releases. TII’s Falcon models from Abu Dhabi. And from India, Sarvam’s models and AI4Bharat’s IndicTrans2 translation models, which I mention because the language coverage matters for work here; check the current license on each before relying on them.
The families, by task
General chat and instruction following
Any instruct model from Qwen3, Llama 3.3, Gemma 3 or Mistral Small will serve a general assistant well. At 7B to 9B choose Qwen3 or Gemma 3 for quality per gigabyte; at 24B to 32B the Gemma 3 27B, Qwen3 32B and Mistral Small are close; at 70B and above Llama 3.3 and the large Qwen3 and DeepSeek models lead.
Reasoning
Models trained to think before answering: DeepSeek R1 and its distills, Qwen3 in thinking mode, gpt-oss with reasoning effort turned up, Magistral, and Phi-4-reasoning at the small end. They are slower and more expensive per answer, so use them for the problems that need them and route simple requests elsewhere.
Code
Qwen3-Coder is my first pick, followed by Devstral for agentic editing and the DeepSeek coder models. StarCoder2 remains useful under its RAIL license for completion. Codestral is strong but its non-production license rules it out for most deployments. The coding assistants that use these models are covered in chapter 6.
Embeddings and reranking
For retrieval, nomic-embed-text (Apache-2.0, with open training data), the BGE and E5 families (MIT) and the Qwen3 embedding models cover almost every need; the Qwen3 and BGE rerankers improve results further. How they fit into a pipeline is in my private RAG guide.
Vision
Qwen’s vision-language models, Gemma 3, the Llama vision models and Mistral’s Pixtral handle image understanding; Moondream is the small, fast option for simple tasks. Licenses follow the parent family.
Speech
OpenAI’s Whisper (MIT, 2022) is still the default for transcription, usually through whisper.cpp or faster-whisper; NVIDIA’s Parakeet models are faster for English. Open text-to-speech has improved quickly and is worth a fresh look whenever you need it.
Sizes and quantization, explained once
Quantization stores a model’s weights in fewer bits than they were trained in, cutting memory and speeding up inference at a small cost in quality. A 16-bit weight takes two bytes; at 4 bits it takes half a byte. The rough memory rule is parameters times bits divided by eight, plus room for the context. So an 8B model needs about 16 GB at 16-bit, 8 GB at 8-bit and 4 to 5 GB at 4-bit, and a 70B model needs roughly 40 GB at 4-bit, which is why 70B is a two-GPU or workstation model and 8B runs on a laptop.
The formats you will meet: GGUF is llama.cpp’s file format, introduced in August 2023, and its quantization names encode the scheme, with Q4_K_M the usual balance point, Q5_K_M and Q6_K closer to full quality, Q8_0 nearly lossless, and the IQ variants for very small files. Ollama pulls Q4_K_M by default. AWQ and GPTQ are GPU-oriented 4-bit formats used by vLLM and similar servers, and FP8 is the choice on recent data-center GPUs. What you lose at 4-bit is small for chat and larger for exact tasks such as code and mathematics, so test at the quantization you will deploy, not at full precision.
To know which of these fits your machine, run AILane; it matches your CPU, RAM and GPU memory against the families above and tells you what will run and at what quantization.
How to choose
- License first. Rule out anything you cannot ship before you look at quality. Chapter 3 tells you how.
- Task second. Pick the family that is strongest at your task, not the one with the best headline score.
- Hardware third. The best model you can run well beats the best model you can barely run. Size to your memory at the quantization you will actually use.
- Context length. Test at the lengths your documents really have; the number on the model card is a ceiling, not a promise.
- Ecosystem. Prefer families with a release cadence, tooling support in your runtime, and fine-tunes you can learn from.
- Your own evaluation. Fifty real examples from your workload, scored the same way for every candidate, will tell you more than any leaderboard. Chapter 5 lists the evaluation tools.
Change log
- September 2026: first published.
Next steps
Continue to chapter 5, The Open Source AI Toolbox, for the runtimes and frameworks that serve these models. To run your first one this afternoon, follow my Ollama guide, and check your hardware with AILane first.