Open Source

Plot twist: Stability AI is not dead and they just dropped StableLM 4 under Apache 2.0. If you've been living under a rock — and honestly with their news cycle you couldn't blame you — there are actual usable weights on HuggingFace today.

The Setup

StableLM 4 ships in 3B, 7B and 12B. The 7B is the obvious daily driver, runs at Q4 on an M4 Mac without breaking a sweat, and Ollama picked it up within 24 hours. Behold the two-command install:

ollama pull stablelm:7b
ollama run stablelm:7b "summarise this changelog in 3 bullets"

# or grab it raw from HF
huggingface-cli download stabilityai/stablelm-4-7b-chat \
  --local-dir ./models/stablelm-4-7b

The Money Pattern

The instruct tuning is shockingly good for a 7B that came out of a near-dead company. It actually follows system prompts and the chat template is sane. I dropped it into a python script for summarising Pipedrive deal notes and the output was usable without post-processing.

from transformers import pipeline

pipe = pipeline(
    "text-generation",
    model="stabilityai/stablelm-4-7b-chat",
    device_map="auto",
    torch_dtype="auto",
)

prompt = [
    {"role": "system", "content": "Summarise sales notes in 3 bullet points."},
    {"role": "user", "content": "Customer called twice, asked about hail repair quote, wants timeline."},
]

out = pipe(prompt, max_new_tokens=200, do_sample=False)
print(out[0]["generated_text"][-1]["content"])

The Catch

It still trails Qwen 3 and Llama 3.3 on most benches. Stability's funding situation is, charitably, rocky — there's no guarantee StableLM 5 ever ships. And the 12B has weird repetition issues above 4k context that smell like training-data hygiene problems.

The Verdict

If you've got an open-weight habit and want a non-Llama, non-Qwen option to keep your stack diverse, StableLM 4 7B is genuinely worth a pull. Don't bet a production line on it given the company's runway, but for self-hosted experiments and side projects it deserves a slot on the M4.