Muse Ember: Meta's Smallest Model, Price, Benchmarks, Open Weights?

Today is Friday, November 6, 2026. Extract the person's name, their company, and the new meeting date from this message. Reply with only a JSON object with the keys "name", "company" and "date" (date as YYYY-MM-DD). No other text.

Message: "Hi, Dana Whitfield here (Northwind Labs). Something came up, so could we push our call to next Tuesday instead? Thanks!"

{"name": "Dana Whitfield", "company": "Northwind Labs", "date": "2026-11-10"}

A shop sells pens in packs of 4 for $3, packs of 7 for $5, or single pens for $1. What is the lowest possible cost for exactly 30 pens? Put only the dollar amount on the last line, like: $9

$22

Write a Python function calc(expr) that evaluates an integer arithmetic expression string with +, -, *, /, parentheses, unary minus and spaces. Division truncates toward zero (so "-7 / 2" is -3). Do not use eval, exec or any parsing library. Reply with only the code in one ```python block.

```python
def calc(expr):
    n = len(expr)
    pos = 0

    def parse_expression():
        nonlocal pos
        value = parse_term()
        while True:
            while pos < n and expr[pos].isspace():
                pos += 1
            if pos < n and expr[pos] in '+-':
                op = expr[pos]
                pos += 1
                rhs = parse_term()
                value = value + rhs if op == '+' else value - rhs
            else:
                return value

    def parse_term():
        nonlocal pos
        value = parse_factor()
        while True:
            while pos < n and expr[pos].isspace():
                pos += 1
            if pos < n and expr[pos] in '*/':
                op = expr[pos]
                pos += 1
                rhs = parse_factor()
                if op == '*':
                    value *= rhs
                else:
                    q = abs(value) // abs(rhs)
                    value = -q if (value < 0) != (rhs < 0) else q
            else:
                return value

    def parse_factor():
        nonlocal pos
        while pos < n and expr[pos].isspace():
            pos += 1
        if pos < n and expr[pos] == '-':
         

A server log of about 950 lines (~30K tokens) with one note hidden 70% of the way in: "NOTE from ops: the vault code for Branch 7 is 4196." The question: "What is the vault code for Branch 7? Reply with only the code."

4196

Tool available: get_weather(city, unit: celsius | fahrenheit). Question: "Is it warmer right now in Lisbon or in Oslo? Use Celsius."

get_weather({"city":"Lisbon","unit":"celsius"})

Write a product description for a solar-powered camping lantern for a US online store. Rules: exactly 3 sentences; no more than 40 words; include the exact phrases "USB-C" and "12 hours"; do not use the words "perfect" or "ultimate". Reply with only the description.

Illuminate your campsite with this solar-powered lantern, featuring bright LED light and rugged weatherproof design. Recharge via solar panel or USB-C cable for convenient backup power anytime. Enjoy up to 12 hours of reliable glow on a single charge tonight.

What is Muse Ember?

Muse Ember is the smallest and cheapest model in Meta’s Muse family, released on November 9, 2026. Meta built it for “the work agents repeat all day”: tool calls, coding subtasks, routing, extraction and parsing long documents. The idea is that a bigger model plans the job and Muse Ember does the many small steps, cheaply and fast.

Key facts from Meta’s announcement:

• Multimodal input: text, images, audio and video in; text out. OpenRouter also lists PDF documents as input.
• 1M-token context window and up to 512K output tokens.
• Reasoning model with adjustable reasoning effort (Meta’s charts show medium, high, xhigh and max settings).
• It is the first model Meta trained using Muse Wonder Full, its largest model.
• Model ID: muse-ember-1.0. It supports the Responses API.

Muse Ember pricing

For comparison inside Meta’s own lineup, Muse Spark 1.3 costs $1.25 / $4.25 and Muse Wonder $4 / $20 per million input / output tokens (OpenRouter list prices). So Ember’s input is more than 12 times cheaper than Spark’s.

Meta’s headline number is cost per task: about $0.14 per Artificial Analysis Intelligence Index task, against $2.39 for Muse Spark 1.3, which scores only one point higher (37 vs 36).

Muse Ember benchmarks

All figures below are from Meta’s announcement, with every model at max reasoning effort. Results marked “AA” come from Artificial Analysis; the others are Meta’s own evaluations.

What this means in practice:

• Same overall score as Haiku 5.5, far cheaper per task. On Meta’s chart, Ember at max effort scores 35.8 for about $0.136 per task; Haiku 5.5 at max scores 36.2 for about $0.411.
• Biggest lead on agent workflows (AutomationBench-AA), which is exactly the job Meta designed it for.
• Not a clean sweep. Haiku 5.5 is clearly ahead on FrontierCode 2.0, which judges whether real engineering changes are good enough to merge. If your work is mostly writing production code, test both.

How well do these numbers hold up? So far two have been re-run by third parties and both match closely. See Benchmarks: official vs independent above.

Is Muse Ember open source / open weights?

No. This is one of the most searched questions, so to be clear: as of November 10, 2026, Muse Ember is API-only. There is no Muse Ember repository on Hugging Face, and Meta has not published its size.

If you need an open-weight Meta model, OpenRouter lists two in the Muse family:

• Muse Spark Open: an open-weight model tuned for agentic coding.
• Muse Glimmer 30B: a dense 30B open-weight multimodal model distilled from Muse Spark, aimed at consumer hardware (131K context).

How to use Muse Ember

There is no free consumer chat for Ember mentioned in Meta’s announcement. It is aimed at developers, and you can reach it in five ways:

1. Meta Model API: sign up at dev.meta.ai and call muse-ember-1.0. Meta says the API works with OpenAI-SDK-compatible clients, so you mostly change the base URL, key and model name.
2. OpenRouter: use the model ID meta/muse-ember. Handy if you already route models through OpenRouter or use a coding agent that supports it.
3. Muse Code: Meta’s terminal coding agent, on a subscription.
4. Google Cloud: available as a partner model.
5. Oracle Cloud: Meta says the Muse models are available there as well.

The setup Meta recommends: let Muse Wonder (or Muse Spark) plan, and hand the high-volume steps (file reads, test runs, small fixes, extraction) to Ember as a subagent.

Who should use it?

• Agent builders paying for many small model calls: routing, classification, extraction, tool calls.
• Coding agent users who want a cheap subagent next to a stronger main model.
• Document and media pipelines: one model reads text, images, audio and video with a 1M context.

It is probably not your best pick as the only model for hard, open-ended reasoning or merge-ready feature work. Meta itself positions Muse Wonder for that.

What we don’t know yet

• Parameter count and architecture
• Whether an open-weight version will follow
• Rate limits and any free tier on Meta Model API
• Independent results on more benchmarks beyond Artificial Analysis and Vals AI

We will update this page as Meta and independent testers publish more.