Open Source

Plot twist: the most interesting English-language LLM this month came out of a lab in Hangzhou. MiraFish dropped its English-tuned 13B and it's surprisingly not a cursed translation layer.

The Setup

MiraFish-EN is a 13B dense model, Apache 2.0, native multilingual training. The English tune was their side quest. The side quest scored 88% on MMLU and 76% on HumanEval. That's main quest energy.

{`huggingface-cli download mirafish/mirafish-en-13b-chat \\
  --local-dir ./mirafish-en

# or just yeet it through vLLM
vllm serve mirafish/mirafish-en-13b-chat --port 8000`}

The Money Pattern

It uses a standard ChatML template, so any OpenAI-compatible client works out of the box. Drop it behind an OpenAI SDK pointed at localhost and your existing code doesn't change.

{`from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed")

resp = client.chat.completions.create(
    model="mirafish/mirafish-en-13b-chat",
    messages=[
        {"role": "system", "content": "You are a senior TypeScript engineer."},
        {"role": "user", "content": "Refactor this React component to use a custom hook."},
    ],
)
print(resp.choices[0].message.content)`}

The Catch

English is still second-class to Mandarin. Idioms get awkward, and it occasionally drifts into a Mandarin token mid-sentence when prompts get long. Tokenizer is fine, vocab is the issue.

The Verdict

For a permissively-licensed 13B in 2026, MiraFish-EN is wild value. Better than Llama 3 13B on reasoning, faster to fine-tune than Mistral. If your stack is Supabase + a self-hosted inference endpoint, this is the new candidate to swap in. Just don't deploy it to a translation product.