by MetaReleased 9 August 2026
Muse Glimmer 30B is a Meta model you can download and run yourself, meant to power automated helpers on ordinary consumer computers. It takes text and pictures, answers in text, and was trained to copy the larger Muse Spark, so it handles multi-step thinking, tool use and recovery after a failed step, and works in more than 100 languages. It reads about 131 pages at once.
You pay for what each request uses. No subscription, no minimum.
One message, the ordinary use
Every rate
| Meter | ↑ you send | ↓ comes back |
|---|---|---|
| Text tokens | 79,88 DA | 319,52 DA |
| Cached tokens read | 10,65 DA |
per 1M tokens
One company runs this model. OpenDunes sends every request to it. DA per 1M tokens: ↑ what you send · ↓ what comes back.
You pay the price of the provider that served your request. Prices read 15:19 · speed over the last 30 minutes · uptime over 3 days.
Swap the base URL and the key. Your OpenAI code runs unchanged.
import os
from openai import OpenAI
client = OpenAI(
base_url="https://opendunes.com/v1",
api_key=os.environ["OPENDUNES_API_KEY"],
)
resp = client.chat.completions.create(
model="meta/muse-glimmer-30b:batch",
messages=[{"role": "user", "content": "Two lines to welcome a shop in Oran."}],
)
print(resp.choices[0].message.content)Runs this request once on your balance.
Standard OpenAI fields, grouped by what they control. Anything not listed is accepted and ignored.
Shaping the reply
How long, how varied, where to stop.
max_tokenstemperaturetop_ptop_kmin_pstopseedfrequency_penaltypresence_penaltyrepetition_penaltylogit_biasReasoning
Let it think before answering — and see the thinking.
reasoningreasoning_effortinclude_reasoningStructured output
Get JSON you can parse instead of prose.
response_formatstructured_outputsTools
Let the model call your functions.
toolstool_choiceAbout 0,24 DA per message. ↑ 79,88 DA · ↓ 319,52 DA per 1M tokens.
No. You pay per use from your balance, with no subscription. Run it once from Call it in code, at the same price.
Muse Glimmer 30B is a Meta model you can download and run yourself, meant to power automated helpers on ordinary consumer computers. It takes text and pictures, answers in text, and was trained to copy the larger Muse Spark, so it handles multi-step thinking, tool use and recovery after a failed step, and works in more than 100 languages. It reads about 131 pages at once.
Deposit dinars with your Edahabia or CIB card. Every model, Muse Glimmer 30B (batch) included, is paid from that balance. No foreign card, no conversion.
Meta names no languages for Muse Glimmer 30B (batch). Arabic, French and Darja: not stated.
It reads text and images. Up to 131K tokens in one request, about 130 pages. It writes up to 12 000 words at once.
Yes, its weights are public. Open weights on Hugging Face
Meta. Released 9 August 2026.
Muse Image makes and edits pictures from a written description, with optional reference pictures. It thinks before it draws, breaks multi-part requests into steps, and can search the web for factual requests. It writes readable text inside the picture, keeps a style or subject steady across a series, and edits by taking the previous result back with a fresh instruction.
Muse Voice Transcribe 1.0 turns speech into text in a single step. Send a single-channel WAV recording of up to 10 minutes and get the transcript back. It handles push-to-talk, finding where a speaker stops, and separate speakers, and you can steer it toward specialist terms or a language with a hint. It gives no word-by-word timing or confidence scores, so convert recordings in other formats to WAV first.
Muse Spark 1.3 Contributor is the lower-cost tier of Muse Spark 1.3, with the same text, picture, video and PDF input and thousand-page memory, for experiments, learning and early automated and coding projects. It keeps track of information across long tasks and works through conflicting information. For the previous release at this tier, use 1.2 Contributor.
#2 for websites · Muse Spark 1.3 is a Meta model that thinks before it answers, reading text, pictures, video and PDFs and replying in text, with about a thousand pages in mind at once. It is built for long-running automated and coding work: it keeps track of facts across a long task, works through conflicting information and asks you to clarify or confirm when it needs to, keeping its steps concise. Meta does not use your requests to improve its products.
From what its maker and two benchmark sites publish, and where it stands on OpenDunes. Read 7 Nov 2026.
Thinks first
Works through the problem before it answers.
Uses your tools
Can call functions your code provides.
Fills your JSON
Answers in the JSON shape you give it.
Searches the web No.
Can look things up online while it answers.
Cheaper on repeats
A prompt it saw a moment ago costs less the second time.
Open weights
The model's weights are public.