Picking a model used to mean picking a single file. Picking a model family means picking an ecosystem: the quant builds that exist, the licence you agree to, the special-purpose siblings (coder, vision, reasoning), and the people who already wrote the system prompts. The choice is durable in a way the model itself is not, because the family decides what you swap to next quarter when the weights change.
This page reads the Hugging Face API and the Ollama library on the day it was last verified, and tells you which family is worth committing to for the workload you actually have.
What “family” buys you that a single model does not
A model family is a vendor’s continuous product line. It includes the dense and mixture-of-experts checkpoints at multiple parameter counts, the code-tuned and instruct-tuned siblings, the vision and multimodal variants, and the long-context and reasoning rewrites. Qwen ships everything from a 0.8B on-device model to a 397B MoE; Mistral ships the same name in Apache-2.0, Modified-MIT, and non-commercial variants; Google ships Gemma under Apache-2.0 in the 4 generation and under a custom “gemma” licence in the 3 generation. The breadth is the point: you can stay inside one family as the workload moves from 8 GB of VRAM to 80 GB without relearning the licence, the prompt format, or the system-message vocabulary.
The first decision is therefore not “which 32B model” but “which 32B family line will still be shipping variants I can move to when this one retires.” A family that ships nothing between today and next year is a family you are about to be stuck inside.
The Families Worth Naming
Six open-weight lines carry the bulk of the open-weight downloads in 2026: Qwen, Llama, Mistral, Gemma, DeepSeek, and Phi. Two more - Granite and gpt-oss - are worth knowing because they cover specific hardware tiers that the others underserve.
| Family | HF licence for current flagship | Gated | Smallest / largest flagship params | Notable shape |
|---|---|---|---|---|
| Qwen (Alibaba) | Apache-2.0 | No | 0.8B / 397B | Full ladder: dense, MoE, coder, vision, reasoning |
| Llama (Meta) | Custom (“llama3.x” / “Llama 4 Community”) | Manual | 1B / 405B | Gating blocks scripted downloads; MAU threshold in licence |
| Mistral (Mistral AI) | Apache-2.0 for the Small 4 line; Modified MIT with revenue cap on Medium | No | 2B / 128B (dense and MoE) | Variant licence - same vendor, different terms |
| Gemma (Google) | Apache-2.0 in Gemma 4; custom “gemma” in Gemma 3 | Manual | 1B / 31B | Two generations under opposite licences |
| DeepSeek (DeepSeek AI) | Apache-2.0 / MIT | No | 1.3B / 685B MoE | MIT family uses commercial-friendly terms |
| Phi (Microsoft) | MIT | No | 0.5B / 14B | Small-model specialist; reasoning siblings |
| Granite (IBM) | Apache-2.0 | No | 1B / 32B MoE | Enterprise-friendly terms, agent and coder siblings |
| gpt-oss (OpenAI) | Apache-2.0 | No | 20B / 120B MoE | The only OpenAI open-weights line |
The numbers above were re-read from the Hugging Face API on 14 September 2026. The Qwen3.5-9B repository returns safetensors.total of 9.65B, last modified 2 March 2026, under Apache-2.0 with gated: false (HF API). Mistral-Small-4-119B-2603 returns safetensors.total of 119.4B, last modified 15 July 2026, also Apache-2.0 and not gated (HF API). Devstral-Small-2-24B-Instruct-2512 sits inside Mistral at 24B params, Apache-2.0, not gated (HF API). DeepSeek-V3.2 returns 685.4B safetensors total, MIT, not gated, last modified 1 December 2025 (HF API). Phi-4-mini-instruct returns 3.84B, MIT, not gated (HF API). Granite-4.0-h-small returns 32.2B, Apache-2.0, not gated (HF API). OpenAI’s gpt-oss-120b returns 116.8B, Apache-2.0, not gated (HF API).
The Two Decision Axes
Two questions dominate which family is the right answer for a reader who already knows their hardware.
Is the licence compatible with what you intend to ship? Apache-2.0 and MIT are the two licences that travel anywhere - including closed commercial products, on-device redistribution, and most government and healthcare work. The custom licences do not. Meta’s Llama 4 licence requires a separate agreement with Meta “if, on the Llama 4 version release date, the monthly active users of the products or services made available by or for Licensee, or Licensee’s affiliates, is greater than 700 million monthly active users in the preceding calendar month” (Meta Llama 4 LICENSE). Google’s Gemma 3 generation ships under a custom “gemma” licence that requires a separate licence for certain distribution patterns, and the Hugging Face repository for google/gemma-3-4b-it returns gated: "manual" (HF API). Meta’s meta-llama/Llama-3.3-70B-Instruct is gated manual too (HF API). Gating means a human at the vendor approves each download request - it does not block you from building on the weights, but it adds an access step that Apache-2.0 repositories do not.
Is the size ladder broad enough to survive your next hardware change? Most readers buy a single GPU and then upgrade in two to three years. A family that ships a 1B, 4B, 8B, 30B, and 70B variant lets you reuse the same prompts, the same Ollama Modelfile, and the same evaluation harness across every upgrade. A family that ships one model at one size forces a rewrite when you move off it. Qwen, Mistral, Phi, and Granite all run a full ladder; DeepSeek and gpt-oss are MoE-only at the top end and dense-only at the bottom; Llama and Gemma publish broader ladders than their flagships suggest but with the gating wrinkle.
What Each Family Is Best At
The “best at” labels below come from how each family is positioned by its maintainers and the workload shapes where their weight sizes land.
- Qwen is the broadest ladder, with Apache-2.0 across the line. The 0.8B-9B band runs on phones and small GPUs, the 30B-35B MoE tier fits one consumer card, and the 122B-397B tier covers workstation-class multi-GPU. The Coder and VL siblings are first-class products, not afterthoughts. The default pick if you want one family to cover chat, code, vision, and a reasoning sibling.
- Llama is the most-downloaded brand on Hugging Face but the gating is real and the licence carries the 700M MAU clause. If the gating friction is acceptable and you are below the MAU threshold, the Llama 3.3 70B Instruct remains a strong chat default at 43 GB on Ollama (Ollama).
- Mistral has the best small-model specialists. Devstral-Small-2-24B-Instruct-2512 is purpose-built for agentic coding on a 24 GB card, and Mistral-Small-4-119B-2603 is a 119B MoE with Apache-2.0 and 6.5B active parameters. Watch the licence per model: Mistral Medium 3.5 128B is a Modified MIT with a 20M-dollar revenue cap that does not appear on the Small line.
- Gemma is the cleanest choice for an enterprise reader who trusts Google but needs Apache-2.0 - in the Gemma 4 generation only. Gemma 3 is gated manual with a custom licence; Gemma 4 is ungated Apache-2.0. Two Google repositories, opposite answers.
- DeepSeek is the right pick if the goal is a reasoning model at the smallest active-parameter budget. DeepSeek-V3.2 is MIT, not gated, and 685.4B safetensors total. The family ships reasoning and coder siblings.
- Phi is the small-model specialist. Phi-4-mini-instruct at 3.84B parameters, MIT, not gated, runs on a phone and beats much larger dense models on focused tasks. Use it when you need a model that is willing to be wrong politely under 4 GB.
- Granite is the IBM enterprise line under Apache-2.0. Granite-4.0-h-small is a 32B MoE with 9B active; the family ships coder, agent, and guardrail siblings. Choose it if you want one vendor’s name across chat, code, and safety-tuned variants.
- gpt-oss is the only OpenAI open-weight line. The 20B tag at 14 GB on Ollama (Ollama) fits a workstation, and the 120B tag at 65 GB is Apache-2.0. Use it if you specifically want a model family that inherits OpenAI’s tokenizer and prompt conventions.
The Licence Traps That Send You Back to Negotiate
Two patterns are misread more often than any others.
Same vendor, two licences. Mistral ships Mistral-Small-4-119B-2603 under Apache-2.0 and Mistral-Medium-3.5-128B under a Modified MIT with a revenue cap. Google ships Gemma 4 under Apache-2.0 and Gemma 3 under a custom licence. Alibaba ships Qwen3.5 under Apache-2.0 and the Qwen2.5 line under a custom “Qwen” licence that includes a 100 million MAU clause on the 72B model. The brand name is not the licence; the specific repository is the licence. Always read the repository, not the press release.
Gated repositories look ungated from Ollama. google/gemma-3-4b-it and meta-llama/Llama-3.3-70B-Instruct are both gated: "manual" on Hugging Face, which means a human at the vendor approves the request before the weights download. Ollama pulls through its own proxy and does not always surface the gating, so a recommendation sized from an Ollama tag can put a reader into an access request they were not warned about. The licence guide keeps the per-repo trap list current.
The Second Family Problem
One family is rarely enough. A realistic stack uses a chat model from one family, a coder model from another (Devstral and Qwen-Coder are stronger than the chat siblings at tool calls), an embedding model (covered separately), and possibly a small Phi or Qwen3 1.7B on-device model for offline phone use. The VRAM tier index starts from a card and walks down to which models fit; the capability ranking ranks the same families by what they can do at the top of the ladder. Use this page to pick the family line, then use those pages to pick the specific checkpoint inside it.
Bottom Line
Commit to Qwen if you want one Apache-2.0 family that covers every size from 0.8B to 397B and every workload from chat to vision. Commit to Mistral if your priority is the strongest small-model specialists under permissive terms and you read the licence per repository. Commit to Llama only if you accept the gating and the 700M MAU clause. Commit to Gemma 4 - not Gemma 3 - if you want a Google-backed Apache-2.0 family. Commit to DeepSeek for reasoning at the smallest active-parameter cost. Commit to Phi for the smallest models worth shipping. Commit to Granite for an IBM-signed enterprise line. Commit to gpt-oss if OpenAI’s prompt conventions matter more than the model’s brand.
Two families is normal; one is the start of a stack, three is the ceiling. The Hugging Face API returns the truth per repository - licence, gating, parameter count, last-modified date - so the rule that prevents getting stuck is the same in every case: read the repository, not the press release, and choose the family whose ladder covers your hardware for the next two upgrades.