Primary use stated or inferred cautiously from official documentation.
META / meta-models/Muse-Spark-Open-816B-A42B
Muse Spark Open
Meta multimodal model released on October 15, 2026, with 816 billion total and 42 billion active parameters. It accepts text, images, audio and video and returns text. It has a combined one-million-token context and BF16 and NVFP4 variants.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | meta-models/Muse-Spark-Open-816B-A42B |
|---|---|
| Model type | Multimodal · Code · Text and reasoning |
| Access model | Restricted weights |
| License | Muse Model License |
| Deployment | Self-hosted or third-party providers |
| Release | 15 OCT 2026 |
| Knowledge cutoff | Not published |
| Entrance | Text · Image · Audio · Video |
| Exit | Text |
| Context window | 1,000,000 tokens |
| maximum output | Combined context: 1,000,000 tokens |
| Reasoning | low / medium / high / xhigh |
| Published tools | Features |
| Structured Outings | Not published |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · SGLang · vLLM · TensorRT-LLM |
Coding, reasoning and tool-using agents. Reasoning effort supports low, medium, high and xhigh; the default is xhigh. The guide recommends temperature 1 and top-p 1. Weights require access approval and large-scale infrastructure.
The Muse license is neither MIT nor Llama: it requires attribution and specific conditions for model-serving businesses. No weights were downloaded or terms accepted. BF16 and NVFP4 are formats of the same model, not two independent launches.
03 / CAPABILITIES
What it can do
Orientation
Coding, reasoning and tool-using agents. Reasoning effort supports low, medium, high and xhigh; the default is xhigh. The guide recommends temperature 1 and top-p 1. Weights require access approval and large-scale infrastructure.
Context
1,000,000 tokens · Combined context: 1,000,000 tokens
Tools and integration
Features
Access
Hugging Face · approved access
Input modalities
TextImageAudioVideo
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | Depends on the provider | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Depends on the provider | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
No single tariff has been verified: hosted APIs belong to third-party providers. The guide states about 700 GB of GPU memory for NVFP4 and 2 TB for BF16. Downloads do not remove infrastructure, operations or safety-review costs.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
Meta-reported results with xhigh effort, temperature 1, top-p 1 and top-k disabled; not reproduced by Inferama. Preserve each test harness. Tool, coding and video metrics are not interchangeable and do not guarantee results for your own agent.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| JobBench | 61,8 | Higher is better | View source ↗ |
| MCP-Atlas | 89,2 | Higher is better | View source ↗ |
| OSWorld Verified | 80,4 | Higher is better | View source ↗ |
| Toolathlon Verified | 72,8 | Higher is better | View source ↗ |
| GPQA Diamond | 89,0 | Higher is better | View source ↗ |
| DeepSWE 1.1 (Preferred harness) | 60 | Higher is better | View source ↗ |
| Terminal-Bench 2.1 (Terminus 2) | 76,4 | Higher is better | View source ↗ |
| ZeroBench (con herramientas) | 50,0 | Higher is better | View source ↗ |
| IntPhys2 | 68,3 | Higher is better | View source ↗ |
| VideoMMEv2 | 61,5 | Higher is better | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
BF16 / NVFP4
The Muse license is neither MIT nor Llama: it requires attribution and specific conditions for model-serving businesses. No weights were downloaded or terms accepted. BF16 and NVFP4 are formats of the same model, not two independent launches.
NVIDIA DFlash2
Auxiliary speculative-decoding checkpoint; requires the target model and is not a standalone assistant.
07 / SOURCES