The open-weight family tree

A handful of families cover nearly all local use, and each ships as a ladder of sizes and post-training flavors. Knowing the tree tells you what any random Hugging Face fine-tune is actually built on — and which base to reach for first. All of these are open-weight; what that does and doesn't get you is covered in open vs closed.

The major families

Honorable mentions: GLM (Zhipu) and Yi (01.AI) — strong bases with long-context variants; OLMo (AllenAI) — fully open: training data, code, and checkpoints all published, the reference point when "open" must mean more than weights.

Sizes and license terms drift per release — read the model card before shipping anything commercial.

Family notes

• Qwen — the default answer for multilingual local use. The size ladder is unusually fine-grained (0.6/1.7/4/8/14/32B dense, plus 30B-A3B and 235B-A22B MoE), so every VRAM tier has a Qwen. The Qwen3 generation added per-prompt "thinking" toggles on most sizes.
• Llama — less the quality frontier than the compatibility frontier: every tool, tutorial, and fine-tune supports it first. Llama 4 moved to MoE, so the dense 3.x line remains the daily driver in many local setups.
• Gemma — consistently strong for its file size, which matters when you're Q4-bound on a laptop. Gemma 3's vision input makes it the easy multimodal default locally.
• Mistral — pioneer of the usable open 7B and of open MoE (Mixtral 8x7B). Recent flagships lean closed; the open 7Bs and Small (22–24B) variants remain workhorses.
• Phi — trained heavily on curated "textbook-quality" synthetic data; punches above its weight on math and logic, thinner on world knowledge. A good small-agent brain.
• DeepSeek — V3 and R1 pushed open reasoning capability forward; the full models need serious hardware, but the R1-distill family (Qwen/Llama bases trained on R1 reasoning traces) runs chain-of-thought locally from 1.5B up.

Generations at a glance

Family trees move fast; generations age like milk. Rough lineage, oldest to newest covered here:

The useful rule: within a family, a newer generation at smaller size usually matches an older generation at larger size. A current 8B routinely beats a two-generation-old 30B at most tasks — which is why the right instinct is "newest model that fits," not "biggest model that fits."

Reading model names: base vs instruct vs reasoning

Every family ships several post-training flavors; the name is the decoder ring:

A base model answers by continuing your text: say hello and you may get a forum thread, an FAQ, or a Wikipedia-style paragraph about greetings. Instruct models behave like assistants. Everything else in this guide assumes instruct or reasoning checkpoints — running a base model and concluding "local models are dumb" is a classic first-week mistake.

MoE: total vs active parameters

Mixture-of-experts models store many more weights than any one token uses. Each layer's feed-forward block splits into expert sub-networks; a small router picks the top few experts per token, per layer. "235B-A22B" means 235B parameters stored, ~22B active per token. Two costs come apart:

memory for weights  ∝ total params    (all experts must sit in RAM/VRAM)
compute per token   ∝ active params   (only the routed experts run)

So an MoE buys big-model quality at small-model speed — if you can hold the weights. That is the entire local story of Qwen3-30B-A3B: ~30B of stored capability for machines with 24–32 GB, at ~3B decode cost, out-scoring same-size dense models while generating faster. The mechanics are in speed.

Same memory budget, different shape:

MoE caveats: quality per active parameter is lower than a dense model's, routing adds overhead at batch size 1, and routers are sensitive to sloppy quantization — prefer community-vetted quant recipes for MoE checkpoints over hand-rolled ones.

Korean and multilingual

Rough community consensus for Korean, as of this writing: Qwen first (strong Korean across the size ladder), Gemma second (clean Korean prose, occasionally stiff honorifics), English-first families a visible third (particle errors, calques from English). Test directly — honorifics (존댓말/반말), particle choice, register consistency — with the protocol in choosing.

The fine-tune jungle

Most repos on Hugging Face are not from the original lab. Layers of culture accumulate on top:

• Merges — weight-averaging (TIES, DARE, SLERP) of several fine-tunes. Sometimes genuinely better than any parent; often a coin flip that evals worse.
• Abliterated — refusal directions surgically removed. Useful for fiction and research; usually costs some general capability. Mind the license and your jurisdiction.
• "Slop" fine-tunes — a quick LoRA on synthetic data chasing arena vibes, published under an escalating name. Quality varies wildly, in both directions.
• Community quants — bartowski, mradermacher, and peers re-quantize upstream without touching the weights. These are safe; it's behavior-modifying fine-tunes that need vetting.

When you open a fine-tune's card, look for the named base model, evals from a harness you recognize (not screenshots), and a described dataset — "private dataset" is a yellow flag.

Default policy: prefer the upstream lab release unless a fine-tune has independent, reproduced evidence of being better. How to weigh that against benchmarks and community signal: choosing.

Related: Choosing a model · Open vs closed · Quantization · Model formats