Meta Releases Muse Spark Open: an 816B Multimodal Model With 1M Context

Meta released two Muse Spark Open checkpoints on October 15

Meta released Muse Spark Open on October 15, 2026, with two official repository IDs: meta-models/Muse-Spark-Open-816B-A42B for the BF16 checkpoint and meta-models/Muse-Spark-Open-816B-A42B-NVFP4 for the NVIDIA-oriented quantized checkpoint. The live records were public but manually gated, meaning a prospective user could inspect metadata and terms but had to submit the requested identity and organization information before downloading restricted files.

This is an artifact release, not merely a promise of future weights. Both repositories expose configuration, tokenizer and sharded-weight inventories, and Hugging Face published runnable integration guidance for Transformers, llama.cpp, SGLang and vLLM. Access remains conditional, however: a visible repository and a deploy button do not prove that a particular organization has gate approval, capacity, regional availability or a production support contract.

One model spans four input modalities and text output

The launch guide describes Muse Spark Open as an approximately 816B-parameter mixture-of-experts model with roughly 42B parameters active for each token. Each MoE block contains 256 experts and routes a token through eight. A separate 1.8B-parameter perception encoder handles images and video, while an audio encoder emits one token every 80 milliseconds. Inputs can combine text, images, audio and video; generated output is text.

Meta positions the model for agentic coding, including reading repositories, editing and debugging code, calling tools and working with screenshots or screen recordings. Those are intended-use and training claims, not evidence that the model safely completes a specific production workflow. Teams still need repository-level tasks, tool-call validation, multimodal adversarial cases and human review at the exact autonomy boundary they intend to permit.

The 1M-token window is a capacity claim with a large systems bill

The configuration sets max_position_embeddings to 1,048,576, shared by text and media tokens. Hugging Face estimates a full-window BF16 key-value cache at about 266 GB and documents default video sampling at two frames per second with a 96-frame cap. Media resolution, frame sampling, audio duration, prompt text, reasoning and output therefore compete for context and prefill resources.

A million-token ceiling does not establish million-token retrieval accuracy, evidence ordering, citation fidelity or acceptable latency. Evaluate at several lengths and evidence positions, including contradictory documents and irrelevant media. Record prefill time, peak memory, time to first token, decode rate, tool accuracy and answer support rather than checking only whether the request fits.

42B active parameters still requires storing an 816B-class model

Only eight experts execute for each token, but all experts must remain available. Hugging Face describes the BF16 safetensors checkpoint as about 1.5 TB and reports a Transformers configuration across four nodes of eight H200 GPUs. The API metadata reports about 822.0B parameters for the BF16 artifact. This is not a conventional 42B model from a memory, transfer or orchestration perspective.

The NVFP4 repository is the more compact official route. Hugging Face says it targets NVIDIA Blackwell and documents a single B200-node configuration. Its hosted-endpoint recipe lists eight RTX Pro 6000 98 GB GPUs for 131,072-token context or eight H200 141 GB GPUs for 1,048,576 tokens. These are launch-partner recipes, not AccessAllGPT measurements or guarantees that every workload will fit with the same concurrency and latency.

The license permits broad use but adds material conditions

The repository labels the artifact with the custom Muse model license rather than Apache, MIT or another standard permissive identifier. The displayed terms grant rights to use, copy, modify, fine-tune, distill and distribute, subject to conditions. Copies and derivatives must retain the copyright, license and incorporated usage policy, and products or documentation using the software must prominently display “Built with Muse Spark Open.”

A separate clause applies when a licensee or affiliate offers models through an API or similar model-as-a-service interface and aggregate gross revenue exceeds $20 million during a consecutive twelve-month period: the licensee must notify Meta and enter a separate written agreement before offering Muse Spark Open or derivatives in that business. The clause excludes specified internal use, research and product-feature use. This summary is not legal advice; organizations should review the full license and usage policy against their exact distribution and service model.

SageMaker adds a managed route, not a serverless API

AWS says Muse Spark Open is available in SageMaker JumpStart in US East (N. Virginia) and US East (Ohio). Its JumpStart IDs are meta-vlm-muse-spark-open-816b-a42b for BF16 and meta-vlm-muse-spark-open-816b-a42b-nvfp4 for NVFP4. The documented pairings are ml.p6-b300.48xlarge for BF16 and ml.p5e.48xlarge for NVFP4. Deployment still requires endpoint quota and EULA acceptance; AWS says provisioning typically takes 15–20 minutes, which is a vendor estimate rather than an observed AccessAllGPT result.

Each route creates a real-time endpoint in the customer’s AWS account. AWS documents OpenAI-compatible messages, streaming, a separate reasoning_content field, optional VPC attachment, KMS volume encryption and container network isolation. That is managed infrastructure, not a pay-only-per-token serverless API: AWS warns that the multi-GPU instance is billed for as long as the endpoint runs, even when it serves no requests. The launch post points to general SageMaker pricing rather than publishing a Muse-specific hourly or token rate, retention rule, quota, latency target or service level.

Pricing the endpoint requires a dated regional quote

NVFP4 needs less weight memory than BF16 and AWS calls it the variant to try first when sizing for cost, but lower precision does not prove lower cost per accepted task. Quality regressions, longer reasoning, retries or extra review can erase an infrastructure saving. Conversely, keeping either large endpoint warm at low utilization can dominate the bill before token volume matters.

Capture the exact Region, instance price, model version, endpoint count, uptime schedule, autoscaling behavior, storage, data transfer, CloudWatch, media processing and support tier. Then measure accepted-task throughput and include reasoning tokens, tool calls and failures. Delete evaluation endpoints promptly; the AWS guide includes cleanup for the model, endpoint and endpoint configuration because stopping requests alone does not stop endpoint charges.

What builders should do next

Start with a paper evaluation before requesting production capacity. Have legal review the Muse license and usage policy, identify whether the product is model-as-a-service, and assign an owner for attribution and derivative-model obligations. In parallel, define a compact evaluation set for code editing, multimodal understanding, long-context retrieval and tool calling, including destructive-action and prompt-injection cases.

Proceed to a constrained trial only if gate access, license fit and a credible hardware or hosted route are all resolved. Compare BF16 or NVFP4 against a smaller baseline on accepted-task rate, unsupported claims, intervention rate, latency, peak memory and total cost. Wait if managed pricing or gate approval is unresolved. Reject the deployment if required attribution, service restrictions, infrastructure, text-only output or measured reliability conflicts with the product contract.

Disclosures

AccessAllGPT received no model access, weights, API credits, AWS credits, hardware, data, benchmark logs, briefing, review, payment or compensation for this article. Meta, Hugging Face and Amazon Web Services did not sponsor, review or endorse it. AccessAllGPT did not test, score or rank Muse Spark Open or its hosting routes. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Meta, Hugging Face or Amazon Web Services. Publication-wide relationships are listed on the disclosures page.