On-prem models / Meta
Run Muse Spark Open 816B-A42B on-prem
Meta Superintelligence Labs' open-weight model released October 15, 2026, with 256 experts (8 active) and a 1.8B perception encoder. It takes text, image, video and audio input and is built for coding and agentic work. Reasoning effort is configurable from low to xhigh.
Memory needed
| Precision | Weights | KV per request, 8K | KV per request, 32K | KV per request, 128K |
|---|---|---|---|---|
| BF16 (full precision) | 1632 GB | 0.92 GB | 2.53 GB | 8.98 GB |
| FP8 | 816 GB | 0.46 GB | 1.27 GB | 4.49 GB |
| INT4 (AWQ / GPTQ) | 457 GB | 0.46 GB | 1.27 GB | 4.49 GB |
| GGUF Q4_K_M | 490 GB | 0.46 GB | 1.27 GB | 4.49 GB |
KV cache from the published config: 62 layers, 16 KV heads x 64 dims, 46 sliding-window layers (2048 tokens). 16-bit cache for BF16 rows, 8-bit otherwise.
Fit for ERP and manufacturing work
Technically the most capable option here for complex ERP agents, multimodal document extraction and multilingual plants. In practice the hardware cost limits it to large enterprises, and the licence's ban on operating heavy machinery means it should not drive shop-floor control; back-office ERP use needs legal sign-off.
Strengths
- Frontier-class open model per Meta's card (for example 89.0 on GPQA Diamond)
- Native multimodal input including documents, screenshots, audio and video
- Long-horizon tool use and coding agents
- Official NVFP4 checkpoint
Watch out for
- Needs at least 700GB of VRAM at NVFP4 and 2TB at BF16 per Meta
- Usage Policy prohibits operation of critical infrastructure and heavy machinery, which needs legal review for plant use
- Gated download and branding requirement
- Default xhigh reasoning effort is slow and token-heavy
GPUs that run it
One user, 8K context, 4-bit
| Hardware | Count | Memory used | Tok/s per user |
|---|---|---|---|
| AMD Instinct MI325X 256GB | 2 | 461 / 460 GB | 131 |
| Apple Mac Studio M3 Ultra 512GB | 1 | 459 / 460 GB | 15 |
| Apple Mac Studio M5 Ultra 512GB | 1 | 459 / 460 GB | 22 |
| NVIDIA B300 (Blackwell Ultra) 270GB | 2 | 461 / 486 GB | 156 |
| NVIDIA H200 SXM 141GB | 4 | 464 / 507 GB | 138 |
| NVIDIA H200 NVL 141GB | 4 | 464 / 507 GB | 138 |
| AMD Instinct MI355X 288GB | 2 | 461 / 518 GB | 156 |
| NVIDIA A100 80GB PCIe | 8 | 470 / 576 GB | 105 |
20 people at once, 16K context, FP8
| Hardware | Count | Memory used | Tok/s per user |
|---|---|---|---|
| AMD Instinct MI325X 256GB | 4 | 837 / 921 GB | 104 |
| Intel Gaudi 3 128GB (HL-325L OAM) | 8 | 843 / 921 GB | 100 |
| NVIDIA B300 (Blackwell Ultra) 270GB | 4 | 837 / 972 GB | 122 |
| NVIDIA H200 SXM 141GB | 8 | 843 / 1015 GB | 112 |
| NVIDIA H200 NVL 141GB | 8 | 843 / 1015 GB | 112 |
| AMD Instinct MI355X 288GB | 4 | 837 / 1036 GB | 122 |
| NVIDIA B200 180GB | 8 | 843 / 1296 GB | 134 |
| AMD Instinct MI300X 192GB | 8 | 843 / 1382 GB | 116 |
Serving and licence
Meta lists SGLang, vLLM, Hugging Face Transformers, TensorRT-LLM, Unsloth and llama.cpp. NVFP4 targets Blackwell; mlx-lm 0.32.2 or newer for Apple silicon. Set reasoning_effort per request.
Muse Model License. Free for commercial use, but products must display Built with Muse Spark Open, model-as-a-service providers above 20 million USD revenue need a separate Meta agreement, and the Usage Policy bars uses including operation of critical infrastructure or heavy machinery.
Design a Muse Spark Open 816B-A42B deployment
Preloaded with Muse Spark Open 816B-A42B. Add your ERP and documents, change GPUs or users, and export the diagram.
Questions
How much VRAM does Muse Spark Open 816B-A42B need?
The weights alone take about 1632 GB in BF16, 816 GB in FP8 and 457 GB in 4-bit. On top of that each request holds a KV cache of about 0.46 GB at 8K tokens of context (FP8 cache), so 20 people generating at once need roughly 9.23 GB more, plus a few GB of runtime overhead.
What is the smallest GPU setup that runs Muse Spark Open 816B-A42B?
For a single user at 8K context in 4-bit, 2 x AMD Instinct MI325X 256GB (256 GB each) fits, at an estimated 131 tokens per second. For a team, size for concurrent requests and context; the builder above does that.
Is Muse Spark Open 816B-A42B allowed for commercial use?
Muse Spark Open 816B-A42B is released under Muse Model License. Free for commercial use, but products must display Built with Muse Spark Open, model-as-a-service providers above 20 million USD revenue need a separate Meta agreement, and the Usage Policy bars uses including operation of critical infrastructure or heavy machinery. Check the licence text on the model card with your legal team before production use.
Which engine should serve Muse Spark Open 816B-A42B on-prem?
Meta lists SGLang, vLLM, Hugging Face Transformers, TensorRT-LLM, Unsloth and llama.cpp. NVFP4 targets Blackwell; mlx-lm 0.32.2 or newer for Apple silicon. Set reasoning_effort per request.