Muse Spark Open 816B: Inside Meta's Largest Open-Weight Model

Meta Superintelligence Labs released Muse Spark Open on October 15, 2026 — an 816-billion-parameter sparse Mixture-of-Experts model designed for agentic coding and long-horizon reasoning. It is the largest open-weight model Meta has released to date, and it arrives with a genuinely interesting set of architectural choices worth understanding in detail.

Architecture: 816B Total, 42B Active

The headline number — 816 billion parameters — is the total weight count across all experts. At inference time, only 42 billion parameters are active per token. The model uses 62 transformer layers, each containing a MoE block with 256 experts, of which 8 are selected per token via learned routing. The hidden dimension is 8192, and the model uses grouped-query attention throughout.

This sparse activation pattern is what makes the model practical to serve. A dense 816B model would require roughly 1.6 TB of memory just for weights in BF16; the MoE structure means the compute per token is closer to a 42B dense model, even though the full parameter space is much larger.

For visual inputs, the model includes a dedicated 1.8-billion-parameter Perception Encoder that processes images and video frames before they enter the main transformer. Audio is handled by a separate encoder. All modalities produce text as output — the model is natively multimodal on the input side but text-only on the output side.

The 1M-Token Context Window

Muse Spark Open supports a 1-million-token context window, enabled by a hybrid attention architecture that alternates between sliding-window attention and full attention across layers. This design keeps the KV cache manageable: full attention at every layer for a 1M-token sequence would require roughly 266 GB of KV cache in BF16, which is prohibitive. The sliding-window layers reduce this significantly for most tokens, while the full-attention layers preserve global context at key positions.

The official deployment guide notes that the KV cache consumes approximately 250 KB per token in BF16 at full context, so practitioners running long-context workloads need to plan memory budgets carefully.

Reasoning Effort Levels

One notable design choice is the model's adjustable reasoning effort. The API accepts a reasoning_effort parameter with four levels: low, medium, high, and xhigh. The default is xhigh, which instructs the model to spend more compute on chain-of-thought before producing a final answer. Lower effort levels reduce latency at the cost of reasoning depth.

This is similar to the effort controls Anthropic introduced in Claude Haiku 5.5, and it reflects a broader trend: rather than shipping a single fixed inference mode, labs are exposing the compute-quality tradeoff directly to developers.

Where It Performs Well — and Where It Doesn't

Meta's benchmarks position Muse Spark Open as a tool-calling and agentic workflow specialist rather than a general-purpose coding leader. On MCP-Atlas, a benchmark for multi-step tool orchestration, it scores 89.2 — ahead of Qwen3.8 (84.5) and GLM 5.3 (84.2). On JobBench, an agent evaluation suite, it scores 61.8, well ahead of alternatives like Inkling (31.1).

The picture is different on raw coding benchmarks. Terminal-Bench 2.1 gives it 76.4, while Kimi K3 leads at 85.0. On DeepSWE 1.1, it scores 60 versus Kimi K3's 68. On the Artificial Analysis Intelligence Index, it sits at 39, behind GLM 5.3 (45) and Kimi K3 (44).

The pattern is consistent: Muse Spark Open is optimized for the orchestration layer — selecting the right tool, managing multi-step state, handling long contexts — rather than for the kind of single-shot code generation that Terminal-Bench measures. Whether that tradeoff is the right one depends on what you're building.

Hardware Requirements and Deployment Options

Running Muse Spark Open requires serious hardware. The BF16 checkpoint needs approximately 2 TB of VRAM, which means a cluster of 32 H200 GPUs or equivalent. The NVFP4 (4-bit) checkpoint brings this down to roughly 530 GB, validated on four NVIDIA Blackwell Ultra GPUs or eight H200s.

For teams without large GPU clusters, Meta has demonstrated the model running on two Apple M5 Ultra Mac Studios (512 GB of unified memory each) connected via Thunderbolt 5 with RDMA, using the MLX framework. AMD Instinct MI355X and MI350X are supported via FP8 and MXFP4 quantization. Unsloth provides 2-bit GGUF quantizations that fit into approximately 300 GB of combined system memory.

Supported inference runtimes at launch include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, and llama.cpp.

Speculative Decoding: DFlash V2 and DSpark

Meta ships two dedicated draft models for speculative decoding, one per inference framework. DFlash V2 is designed for SGLang and proposes blocks of tokens in a single CUDA graph pass; on AMD MI355X hardware, it reportedly achieves 3.7× throughput improvement with an average of 6.4 accepted draft tokens per step. DSpark is designed for vLLM and proposes up to seven tokens per pass, with architecture specifically tuned to remain efficient at 1M-token context lengths by leveraging the model's sliding-window attention layers.

Speculative decoding at this scale is non-trivial — the draft model needs to be fast enough that the overhead of running it doesn't eat the gains from fewer full-model forward passes. Shipping framework-specific draft models rather than a single generic one suggests Meta has tuned these carefully for each runtime's batching and scheduling behavior.

Licensing: Open-Weight, Not Open-Source

Muse Spark Open is released under the Muse Model License, which is worth reading carefully before building on it. The key points:

• Any product using the model must display "Built with Muse Spark Open" prominently.
• Companies offering the model as an API service (Model-as-a-Service) with gross revenue exceeding $20 million in a 12-month period must enter a separate written agreement with Meta.
• Fine-tuning and redistribution are permitted within these constraints.

This is the same general structure Meta used for earlier Llama releases — open enough for most research and product use cases, but with guardrails for large commercial API providers.

Practitioner Takeaways

Muse Spark Open is a serious model for teams building agentic systems that need strong tool orchestration and long-context handling. Its MCP-Atlas lead suggests it has been specifically optimized for the kind of multi-step, tool-heavy workflows that are increasingly common in production AI systems.

The hardware bar is high. Unless you have access to a multi-GPU cluster or a pair of M5 Ultra Mac Studios, you'll be working with quantized versions, which may shift the benchmark picture. The NVFP4 checkpoint is the practical starting point for most teams.

The benchmark gaps on Terminal-Bench and DeepSWE are real and worth factoring in if your use case is primarily single-shot code generation rather than agentic orchestration. For that workload, Kimi K3 or GLM 5.3 may still be the better choice.

What's notable about this release is less the raw parameter count and more the deliberate architectural choices: the hybrid attention for 1M-context efficiency, the framework-specific speculative decoding draft models, and the adjustable reasoning effort levels. These are the kinds of engineering decisions that reflect a model built for deployment, not just for benchmark headlines.