Open Source

Spoiler: 01.ai is back. Yi-2 dropped in 9B, 34B and 100B and the open-weight scene has another serious player. If you've been living under a rock since the first Yi release, this one is genuinely worth booting up.

The Setup

Yi-2 is on HuggingFace, the chat template is documented, and the 9B fits comfortably on an M4 Mac at Q4. The multilingual coverage is the real headline — Chinese, English, Japanese, Korean and a credible long tail. Behold:

{`huggingface-cli download 01-ai/Yi-2-9B-Chat \\
  --local-dir ./models/yi-2-9b

# or via ollama for a 1-command run
ollama pull yi-2:9b
ollama run yi-2:9b "translate this paragraph to Mandarin, keep tone formal"`}

The Money Pattern

The 34B is the one to watch. It punches inside Llama 3.3 70B territory on multilingual reasoning while running on a single 48GB GPU at Q5. The chat template handles tool calls cleanly and works with the standard transformers `apply_chat_template` flow without monkeypatching.

{`from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("01-ai/Yi-2-34B-Chat")
model = AutoModelForCausalLM.from_pretrained(
    "01-ai/Yi-2-34B-Chat",
    torch_dtype="auto",
    device_map="auto",
)

msgs = [
    {"role": "system", "content": "You are a bilingual assistant."},
    {"role": "user", "content": "翻译这段话: deploy to Netlify and wire Supabase auth"},
]
ids = tok.apply_chat_template(msgs, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(ids, max_new_tokens=256)[0]))`}

The Catch

Plot twist — the licensing is not uniform. The 9B is Apache 2.0, the 34B is "Yi Community License" which is permissive but not OSI-Apache, and the 100B has commercial-use clauses you actually need to read. Do not @ me — read the LICENSE file per size before shipping anything.

The Verdict

Yi-2 9B is a legitimate alternative to Llama 3 8B for multilingual workloads. The 34B is the genuine sleeper of 2026 so far. The 100B is too big to be practical for solo devs but earns its place at the top of the leaderboards. Pull the 9B tonight, run it on the M4, and see what you've been missing.