Muse Spark Open 816B-A42B
Muse Spark Open 816B-A42B needs roughly 560.3 GB VRAM at Q4_K_M quantization (1837.8 GB at FP16). 2 GPUs we track can run it fully in VRAM at 8k context.
2 GPUs run this natively · 0 with CPU offload
- Q4_K_M total
- 560.3 GB
- at 8k context
- KV cache, full context
- 69.1 GB
- 16 of 62 layers grow with context, 46 capped at 2,048 tokens
- Inputs
- Text · Image · Video · Audio
- Muse Model License, released October 2026
- Attention pattern
- 46 local : 16 global
- 2,048-token sliding window
- Experts
- 8 of 256
- routed per token, about 42B parameters active
Muse Spark Open 816B-A42B is a Mixture of Experts (MoE) model with 820B total parameters but only 42B active per token developed by Meta. Released 15 October 2026 under the Muse Model License as the open-weights version of Meta's hosted Muse Spark, with extra safety post-training for standalone use; PyTorch's launch post says it shares a base model with the hosted Muse Spark 1.3. Each token is routed to 8 of 256 experts. Meta's headline count is 816B total parameters, against the 820B in the weights file. The 62 layers use a sliding-window pattern, 46 local layers on a 2,048-token window and 16 full-attention layers, with 128 query heads over 16 KV heads at a head dimension of 64. The context window is 1M tokens, and input is text, image, video, and audio through separate encoders.
To run Muse Spark Open 816B-A42B locally: At Q4_K_M the weights alone are 499.4GB and the total is 560.3GB at 8k context, more than one 512GB Apple M5 Ultra can hold: the calculator's best fit there is Q3_K_M at 442.8GB. Meta's recommended Mac setup is two M5 Ultra Mac Studios linked over Thunderbolt 5 and running its NVFP4 checkpoint through MLX, a 1TB pool where this site's calculator (which limits NVFP4 to Blackwell GPUs) fits Q6_K at 755.0GB. On NVIDIA, Meta's own NVFP4 checkpoint is 533GB on disk and totals 598.0GB at 8k context; Meta says four Blackwell Ultra GPUs or eight H200s serve it. As a MoE model, inference speed depends on active parameters (42B) rather than total size.
Meta's own comparison table, run at xhigh reasoning effort, has it ahead of every other model in that table on MCP-Atlas (89.2, against 84.5 for Qwen3.8-2.4T-A95B and 84.2 for GLM 5.3), JobBench (61.8), HLE-Diamond without tools (26.1), and AA Omniscience (25). It trails on GPQA Diamond (89.0, against 93.5 for Qwen3.8-2.4T-A95B and Kimi K3 and 91.7 for GLM 5.3), on Terminal-Bench 2.1 (76.4, against 83.9 for GLM 5.3 and 85.0 for Kimi K3), and on the Artificial Analysis Intelligence Index (39, against 45 for GLM 5.3 and 44 for Kimi K3). Every figure is Meta's own, as reported on the model card at release.
Sixteen full-attention layers carry the whole 1M-token window
The expert count is the familiar part: 256 experts in every layer and 8 chosen per token. The attention stack is the less obvious part. Meta's model card lists 62 layers and 128 query heads over 16 KV heads, and says nothing about a sliding window, but the shipped config.json (and the llama.cpp GGUFs converted from it) marks 46 layers as sliding-window attention on a 2,048-token window and 16 as full attention; those 16 also run without rotary position embeddings. A sliding-window layer only sees the most recent 2,048 tokens however long the conversation gets, while a full-attention layer can see all the way back to the first token. The stack opens on a local layer, then repeats one global layer followed by three local ones fifteen times, and closes on a global layer.
16 of 62 layers keep a KV cache that grows with the full context. 46 more cap their cache at a fixed 2,048-token window, no matter how long the context gets.
Past the first 2,048 tokens the 46 local layers stop growing, and only the 16 full-attention layers keep accumulating, at about 65 KB per token. That makes the full 1,048,576-token window cost 69.1 GB of KV cache at FP16. Hugging Face's launch blog quotes about 266 GB for the same window because it multiplies all 62 layers; that is the no-window upper bound (266.3 GB on the chart below), and what a runtime that kept every layer's cache in full would need.
What the full 1M-token window actually costs: 69.1 GB, not 266 GB
On a model this size the weights dominate: 499.4 GB at Q4_K_M against less than a gigabyte of KV cache at 8k context. Context only starts to matter at the long end, which is where the split earns its keep, since 46 of the 62 layers stop adding cache after 2,048 tokens.
KV cache only, at FP16. Weights and activation overhead sit on top of these figures.
Filling the window costs 69.1 GB of KV cache, against 266.3 GB for the same stack with every layer on full attention, 3.9 times less. At Q4_K_M that puts the full-context total at 636.7 GB rather than 857.5 GB. Short of the very top end the cache is a rounding error next to the weights: 0.9 GB at 8k context and 9.0 GB at 128k, so the quantization you can afford decides what fits far more than how much context you use.
VRAM at each quantization
Calculated at 8k context. Because only part of this model's stack keeps a growing KV cache, longer sessions cost far less extra VRAM than a full-attention model of the same size.
| Quant | Weights | KV cache | Total |
|---|---|---|---|
| FP32 | 3280.0 GB | 0.92 GB | 3674.6 GB |
| BF16 | 1640.0 GB | 0.92 GB | 1837.8 GB |
| FP16 | 1640.0 GB | 0.92 GB | 1837.8 GB |
| Q8_0 | 871.7 GB | 0.92 GB | 977.3 GB |
| Q6_K | 673.2 GB | 0.92 GB | 755.0 GB |
| Q5_K_M | 583.8 GB | 0.92 GB | 654.9 GB |
| Q4_K_Mrec | 499.4 GB | 0.92 GB | 560.3 GB |
| Q3_K_M | 394.4 GB | 0.92 GB | 442.8 GB |
| Q2_K | 312.4 GB | 0.92 GB | 350.9 GB |
| NVFP4cuda | 533.0 GB | 0.92 GB | 598.0 GB |
Shown at 8k context with FP16 KV cache. NVFP4 needs a CUDA GPU to run. Toggle TurboQuant in the calculator to view compressed KV cache numbers.
Benchmarks
Against the closest-sized open model: a split decision
GLM 5.3 (753B total, 40B active) is the nearest peer by size on this site's board and sits in Meta's own comparison table, so it is the fairest reference here. On Meta's numbers Muse Spark Open is ahead on MCP-Atlas and HLE-Diamond, level on JobBench, and behind on the rest, most clearly on Terminal-Bench 2.1 (7.5 points) and DeepSWE (7 points).
Meta's own comparison table on the Hugging Face model card: Muse Spark Open at xhigh reasoning effort, GLM 5.3 at its max setting, both as run and reported by Meta, with each model's preferred harness for DeepSWE. That is why GLM 5.3's Terminal-Bench 2.1 here (83.9) is lower than the 88.2 Z.ai reports for it, which this site lists on GLM 5.3's own page. Most of these suites have no field in this site's benchmark data, and none is independently reproduced.
Neither model dominates on Meta's numbers. Muse Spark Open's edge is tool orchestration and hard-knowledge questions; GLM 5.3's is software engineering and terminal work. For a coding agent the Terminal-Bench and DeepSWE gaps are the rows to weigh; for tool orchestration or research, MCP-Atlas and HLE-Diamond are.
GPUs that run Muse Spark Open 816B-A42B natively (2)
- Apple M5 Ultra (512GB)Q3_K_M · 28.1 t/s
- Apple M3 Ultra (512GB)Q3_K_M · 21.8 t/s
Cloud GPU Rental
Cheapest way to run Muse Spark Open 816B-A42B: 8× A100 80GB at Q4_K_M, ~31 tok/s, from $9.52/hr.
Pay by the hour · no contract · pods start in about a minute.
Rent this setup on RunPod ↗ (+$5 signup credit)Prices as of Oct 13, 2026. Rates change; check RunPod for current pricing.
Affiliate link: CanItRun may earn a commission. Doesn't affect the fit calculation above.
Notes
Meta's open-weights version of Muse Spark, released on 15 October 2026 under the Muse Model License with extra safety post-training for standalone use; PyTorch's launch post says it shares a base model with the hosted Muse Spark 1.3. Sparse mixture of experts with 256 experts and 8 routed per token (about 42B active) and a 1M-token context window. Its 62 layers follow a sliding-window pattern: 46 local layers on a 2,048-token window and 16 full-attention layers, read from the shipped config.json (the model card's own spec table does not mention it), which is why the KV cache stays small next to the weights. Input is text, image, video, and audio through separate encoders that ship as an mmproj file in the GGUFs and are not counted in the parameter total above; output is text. Reasoning effort is selectable (low, medium, high, xhigh), with xhigh as the default and the setting Meta benchmarked. Benchmark figures are Meta's own, run at xhigh against Inkling, Nemotron 3 Ultra, Beam, Mistral Large 4, Qwen3.8-2.4T-A95B, GLM 5.3, and Kimi K3: it leads that table on MCP-Atlas (89.2), JobBench (61.8), and HLE-Diamond without tools (26.1), but trails Qwen3.8, Kimi K3, and GLM 5.3 on GPQA Diamond (89.0 against 93.5, 93.5, and 91.7) and on the Artificial Analysis Intelligence Index (39 against 40, 44, and 45). Real download sizes: the BF16 checkpoint is 1,644 GB, Meta's own NVFP4 checkpoint is about 533 GB (only the experts are 4-bit), and community GGUFs run from 304 GB (Q2_K) through 470 GB (Q4_K_S) to 872 GB (Q8_0). The license allows commercial use with a "Built with Muse Spark Open" notice and requires a separate agreement with Meta for Model-as-a-Service businesses above $20M in annual gross revenue.
Compare Muse Spark Open 816B-A42B with other models
How to run Muse Spark Open 816B-A42B locally
Q4_K_M needs 560.3 GB: needs multiple datacenter-class GPUs (80 GB+ each).
vLLM with Meta's NVFP4 checkpoint (four Blackwell Ultra GPUs or eight H200s)
vllm serve meta-models/Muse-Spark-Open-816B-A42B-NVFP4 \
--tensor-parallel-size 4 \
--enable-expert-parallel \
--max-model-len 1048576 \
--reasoning-parser muse \
--tool-call-parser muse \
--enable-auto-tool-choice
# Meta's full recipe also adds DSpark speculative decoding for speed.
# The weights repo is gated: accept the Muse Model License on Hugging Face first.llama.cpp
# Prebuilt binaries: curl -LsSf https://llama.app/install.sh | sh
llama serve -hf ggml-org/Muse-Spark-Open-816B-A42B-GGUF:Q4_K_S
# Model card sampling: temperature 1.0, top_p 1.0, top-k disabled (the GGUF
# metadata ships top_k 64 instead). Reasoning cannot be turned off; the effort
# levels are low, medium, high, and xhigh (the default), and Meta recommends
# high or xhigh for coding and agent work.Why this quantization? Q4_K_M is the standard 4-bit build: about 499.4 GB of weights, and the real files bracket that figure (470.1 GB for ggml-org's Q4_K_S, the build Meta's own llama.cpp example uses, and 505.5 GB for Unsloth's UD-Q4_K_XL). The total is 560.3 GB at 8k context and 636.7 GB with the full 1M-token window filled, and the window adds less than it might because only 16 of the 62 layers keep a KV cache that grows with context. On Blackwell hardware use Meta's own NVFP4 checkpoint instead (533 GB of weights, 598.0 GB in total at 8k). Going below Q4 costs real quality: Unsloth's own guide treats its UD-IQ2_S build (275.5 GB) as the minimum for agentic work, since smaller quants degrade on general tasks.
Who is Muse Spark Open 816B-A42B for?
Teams with a four-GPU Blackwell Ultra node, eight H200s, or a pair of 512 GB M5 Ultra Mac Studios who want a frontier-class open-weights coding and agent model with text, image, video, and audio input and a 1M-token window, and who can live with the Muse Model License's terms.
Best for
- Long-horizon coding agents that read a whole repository inside the 1M-token window
- Tool-heavy agent workflows: Meta's own comparison table has it ahead on MCP-Atlas (89.2) and JobBench (61.8)
- Multimodal agents that read screenshots, documents, audio, and video
- Private, on-premises deployment of a frontier-class model on a pair of Mac Studios
- LLM-as-a-judge evaluation and synthetic training-data generation, two uses Meta lists for it
Not ideal for
- Any single consumer GPU, or one 512 GB Mac at Q4 or better: Q3_K_M is the ceiling on a single M5 Ultra
- Latency-sensitive chat: reasoning cannot be turned off, only dialed down to low
- Teams that need an OSI-style open license: the Muse Model License requires a "Built with Muse Spark Open" notice and a separate agreement with Meta for Model-as-a-Service businesses above $20M in annual gross revenue
- Anyone who needs independently verified benchmarks, since every score at release is Meta's own
So should you run it?
Only if you already have a multi-machine budget. No single GPU this site tracks holds it, and one 512 GB M5 Ultra tops out at Q3_K_M (442.8 GB at 8k context) or Q2_K (427.3 GB with the full 1M-token window filled). The setups Meta points to fit it comfortably by this site's calculator: two 512 GB M5 Ultras take Q6_K at 755.0 GB, and four B300s take the 533 GB NVFP4 checkpoint at 598.0 GB (674.4 GB with the full window). Treat the speeds as a ceiling. Hugging Face's authors report 18 tokens per second for a llama.cpp 2-bit build on a 512 GB M3 Ultra, where this site's calculator estimates 26.0; their MLX 3-bit build reports about 21 tokens per second, close to the 22.3 estimated here for Q3_K_M. On quality, Meta's own table shows a model that leads on tool use (MCP-Atlas 89.2) and hard knowledge questions (HLE-Diamond 26.1) but trails GLM 5.3 on GPQA Diamond and the coding rows, and every figure is Meta's until an independent leaderboard reproduces it. Reasoning cannot be turned off, only dialed down to low.
Frequently asked questions
- What are the VRAM requirements for Muse Spark Open 816B-A42B?
- Muse Spark Open 816B-A42B requires approximately 560.3 GB of VRAM at Q4_K_M quantization, 977.3 GB at Q8, and 1837.8 GB at FP16. These numbers assume 8k context window; its hybrid attention stack caches far fewer than all layers, so VRAM grows much slower than linearly with context.
- How many parameters does Muse Spark Open 816B-A42B have?
- Muse Spark Open 816B-A42B has 820 billion total parameters, but only 42 billion are active per token thanks to its Mixture of Experts (MoE) architecture. This makes inference significantly faster than the total parameter count suggests.
- Can Muse Spark Open 816B-A42B run on a 16 GB GPU?
- No. At Q4_K_M, Muse Spark Open 816B-A42B needs 560.3 GB of VRAM, more than 16 GB. You will need a multi-GPU server.
- Can Muse Spark Open 816B-A42B run on a 24 GB GPU?
- No. Even at Q4_K_M, Muse Spark Open 816B-A42B needs 560.3 GB. Consider a multi-GPU server with 561 GB+ of combined VRAM.
- What is the smallest quantization for Muse Spark Open 816B-A42B that fits in 24 GB of VRAM?
- Muse Spark Open 816B-A42B cannot fit in 24 GB of VRAM at any standard quantization level. The minimum needed is 350.9 GB at Q2_K.
- What GPU do I need to run Muse Spark Open 816B-A42B locally?
- You need a multi-GPU server. At Q4_K_M, Muse Spark Open 816B-A42B needs 560.3 GB VRAM, more than any single consumer GPU. That's roughly 8x 80 GB datacenter GPUs (H100, A100, or similar) pooled together.
- Can I run Muse Spark Open 816B-A42B in the cloud?
- Yes. The cheapest RunPod setup we found that holds Muse Spark Open 816B-A42B entirely in GPU memory at 8k context is 8× A100 80GB at Q4_K_M, at roughly 30.8 tokens per second and from $9.52 per hour for the whole pod. Prices as of Oct 13, 2026; RunPod rates change, so check their pricing page for the current figure.
- Can I run Muse Spark Open 816B on a Mac Studio?
- Two 512 GB M5 Ultra Mac Studios linked over Thunderbolt 5 is the setup Meta recommends, running its NVFP4 checkpoint through MLX; four 256 GB machines also work. One 512 GB M5 Ultra is not enough for Q4: by this site's calculator its best fit is Q3_K_M at 442.8 GB at 8k context, or Q2_K at 427.3 GB with the full 1M-token window filled. Hugging Face's authors ran a 2-bit llama.cpp build on a 512 GB M3 Ultra at 18 tokens per second, with 304 GB of that machine's memory used by the weights.
- Why does this page say 820B when Meta says 816B?
- Meta's branding is 816B total with 42B active. The language model's tensors in the llama.cpp GGUFs count 820.0B parameters, and the file sizes (so the VRAM math here) follow that: the BF16 GGUF is 1,640 GB, exactly 820B at two bytes each. The gap looks like the roughly 3.3B parameters of separate input and output embeddings, though Meta does not say so. The perception and audio encoders, about 2B parameters, ship as a separate mmproj file and are not counted.
- Is Muse Spark Open the same model as Meta's hosted Muse Spark?
- It is the open-weights version. Meta describes it as the open weights version of Muse Spark, with extra safety post-training for standalone use, and PyTorch's launch post says it shares a base model with the hosted Muse Spark 1.3. The hosted Muse Spark models stay API-only; this page covers only the downloadable weights, which sit behind a gated Hugging Face repo that needs the Muse Model License accepted first.
- How big is Muse Spark Open's KV cache at 1M tokens?
- 69.1 GB at FP16 by this site's calculator, because only 16 of its 62 layers keep a cache that grows with context. Hugging Face's launch blog quotes about 266 GB, which multiplies all 62 layers and ignores the 2,048-token window the model's config sets on the other 46; that is the upper bound for a runtime that caches every layer in full, so check how your engine handles sliding-window layers before budgeting memory around the smaller figure.
- Can I use Muse Spark Open commercially?
- Yes, with conditions. The Muse Model License lets you use, modify, fine-tune, distill, and redistribute the weights, and Meta claims no ownership of outputs. It requires a "Built with Muse Spark Open" notice in products that use the model, and a separate written agreement with Meta before you offer the model or a derivative as a Model-as-a-Service if you and your affiliates exceed $20 million in gross revenue in any 12 months. The Usage Policy also rules out uses such as military and weapons applications, and the model is not intended for anyone under 18.