Muse Glimmer 30B
A 30B multimodal model distilled from Muse Spark and purpose-built for autonomous agentic tasks — reasoning, tool use, and failure recovery in a small model.
Total Parameters
30B
Dense + image encoder
Context Window
131K
Tokens
Intelligence
35
Intelligence index
Quantization
BF16
Open weights
Agentic capability at small-model cost
Muse Glimmer is a 30-billion-parameter language model with a dedicated image encoder, distilled from Muse Spark and purpose-built for autonomous agentic tasks on consumer hardware. It integrates multi-step reasoning, reliable tool use, multimodal understanding, and failure recovery into a small model.
On Doubleword it runs on the same OpenAI-compatible async and batch endpoints as the rest of the catalogue, with a 131K context window and BF16 weights.
30B — Vision + Reasoning
Built for high-volume agent work
Autonomous Agents
Purpose-built for multi-step agentic tasks with reliable tool use and structured outputs at small-model cost.
Multimodal Understanding
A dedicated image encoder handles screenshots, charts, and document pages alongside text in the same request.
Failure Recovery
Distilled from Muse Spark to detect dead ends, self-correct, and retry — critical for long-running async pipelines.
Consumer Hardware Scale
30B dense footprint keeps throughput high and cost low for high-volume evals, labelling, and pipeline work.
Reasoning is controlled by prompt, not parameters
Set reasoning strength in the system prompt as Reasoning strength: <value>. Request parameters are ignored.
| Level | When to use it |
|---|---|
| low | Simple extraction, classification, and formatting where latency matters most. |
| medium | Everyday summarisation, drafting, and light tool use. |
| high | Complex problem solving, coding, and agentic tasks. |
| xhigh | Hardest multi-step reasoning and long agent trajectories. |
Use high or xhigh for complex problem solving, coding, and agentic tasks.
Flexible pricing tiers
Choose the balance of speed and cost for your workload. Prices are per 1M tokens.
| Tier | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| Realtime | $0.11 | $0.32 |
| Async | $0.07 | $0.24 |
| Batch (24h)Cheapest | $0.05 | $0.15 |
Context window natively supported up to 131K tokens. BF16 quantization.
Start building in minutes
Muse Glimmer 30B is served over OpenAI-compatible endpoints — point the standard Python SDK at Doubleword.
from openai import OpenAI
client = OpenAI(
api_key="your-api-key-here",
base_url="https://api.doubleword.ai/v1"
)
response = client.chat.completions.create(
model="meta-models/Muse-Glimmer-30B",
messages=[
{"role": "system", "content": "Reasoning strength: high"},
{"role": "user", "content": "Plan and execute a 3-step research task."},
],
)
print(response.choices[0].message.content)# Async / batch: same model, up to 55% cheaper
with open("batch_requests.jsonl", "rb") as file:
batch_file = client.files.create(file=file, purpose="batch")
batch = client.batches.create(
input_file_id=batch_file.id,
endpoint="/v1/chat/completions",
completion_window="24h",
)
print(f"Batch ID: {batch.id}")Pro tip
For agent trajectories, pin Reasoning strength: high in the system prompt and run the workload on the async tier — same behaviour, roughly a third off realtime pricing.