Muse Spark 1.3 & Glimmer: Meta’s Double Release

1. The Spec Sheet: Two Models, Two Philosophies

Spark 1.3 is Meta Superintelligence Labs iterating on its proprietary line — Spark 1.0 in April 2026, then 1.1, then 1.2 on August 5 with the Muse Code terminal agent, the Meta Model API and OpenRouter. Glimmer is the same lab going the other way: open weights on Hugging Face from day one.

2. What Changed Since the Llama Era

Glimmer is Meta’s first open-weight release since Llama 4, and the license alone tells the story: Apache 2.0 replaces the Llama Community License, so commercial use, modification and redistribution work without the old gates.

The design target changed too. Llama competed on raw capability against GPT and Gemini. Glimmer competes on where it runs — always-on local agents, function calling, LLM-as-a-judge — distilled from the flagship instead of trying to be the flagship.

3. Quickstart: Run Glimmer Locally in 5 Minutes

Every command below is verified against the official Ollama library page and the Hugging Face repo. No flags invented, no steps skipped.

Step 1 — Pull and run with Ollama

One command downloads and starts the model. On Apple Silicon with the MLX engine, use the muse-glimmer:30b-mlx tag instead.

ollama run muse-glimmer

ollama run muse-glimmer:30b-mlx

Step 2 — Talk to it over HTTP

Ollama serves a local API on port 11434. This exact call comes from the official library page:

curl http://localhost:11434/api/chat -d '{
  "model": "muse-glimmer",
  "messages": [{"role": "user", "content": "Hello!"}]
}'

Step 3 — Same model from Python

Prefer code over curl? The official Ollama Python snippet works unchanged:

from ollama import chat

response = chat(
    model="muse-glimmer",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(response.message.content)

4. Honest Limits

Two strong releases, but neither is magic. Here is where each one bites.

5. Verdict

If you build agents that must run anywhere — offline laptop, on-prem box, privacy-sensitive client — Glimmer is the most interesting Meta release since the original Llama. Apache 2.0, one GPU, genuinely good tool use.

If you sell coding output by the token, Spark 1.3 is the workhorse: 1M context, fewer wasted calls, and a contributor tier priced for experimentation. Watch for the promised open Spark weights — that release would merge both lanes.

Conclusion

Meta closed August and opened September with a clear message: the flagship stays in the cloud, the agent moves onto your desk. Spark 1.3 pushes coding efficiency forward; Glimmer hands you 30B of distilled agency under Apache 2.0.

My recommendation stands: Glimmer on your machine for always-on agents, Spark 1.3 in Muse Code for heavy coding. When the open Spark weights land, I will update this guide — that is the release that could change both answers.