Muse Spark Open: how fast is it on 3x A100?
4 workload results across 1 hardware configuration.
Fastest local config
26.3 decode tok/s
on 3x A100 (80GB) + AMD EPYC 7742 64-Core Processor (128c) + 1008GB via llama.cpp (IQ1_S). see full run
Muse Spark Open runs at 25.2 tok/s on three A100 80GB cards using Unsloth's 175.6 GB 1-bit file, roughly matching Apple's unverified ~25 tok/s claim for two Mac Studios.
Setup: Meta Muse Spark Open (816B total / 42B active, 1M context). Unsloth UD-IQ1_S GGUF 175.6 GB (1-bit, the smallest file). llama.cpp 2851b3bc (build 11675). 3x A100 80GB, 16K context, whole model on GPU. Peak VRAM ~168 GiB across the three cards. Run cost $1.52. All decode figures are single runs, not medians. Time to first token and prefill are from suite-v1 chat-long: this model's tokenizer counted 3,227 prompt tokens (suite target 4,096). We measure speed and memory only, not output quality.
Measured workloads
| Workload | Decode | Run |
|---|---|---|
chat-short | 25.2 tok/s | r_fsd9qjmmna0 |
chat-long | 25.5 tok/s | r_fsd9qjmmna0 |
concurrent-decode | 26.3 tok/s | r_fsd9qjmmna0 |
agent-trace | 22.8 tok/s | r_fsd9qjmmna0 |
Prefill on the chat-long prompt (3,227 tokens): 288.7 tok/s. Time to first token: 11.2 s.
| Reported hardware | Model · quant | Backend · workload | Decode (measured) | TTFT (measured) | Run |
|---|---|---|---|---|---|
| A100 (80GB) | Muse-Spark-Open-816B-A42B-UD-00001-of-00005 IQ1_S | llama.cpp 0.7.0-dev (build 11675, commit 2851b3bcb) chat-short · 3x A100 80GB @ 16K context | 25.20tok/s | 2,732ms | r_fsd9qjmmna0 2026-10-16 |
What hardware do you need?
- One MacBook / one 24GB GPU: no. Even this smallest 1-bit file is 175.6 GB on disk; our peak was ~168 GiB across three A100 80GB cards.
- Multi-GPU cloud (~240 GB class): yes for this 1-bit file. Our setup was 3x A100 80GB at 16K context, fully on GPU (r_fsd9qjmmna0).
- Two / four Mac Studios (MLX): Meta/Apple's published path. Apple/MLX claim about 25 tok/s on 2 Studios and about 36 on 4, unverified by us and not in our tables. Our 25.2 tok/s on 3x A100 roughly matches the 2-Studio claim in speed only (different hardware and quant).
- Have a multi-Studio setup? Submit an MLX run with the request button so we can set measured numbers against Apple's claims. See also VRAM fit and the open dataset.
FAQ
Can I run Muse Spark Open locally on one Mac or one GPU?
Not on one MacBook or one 24GB gaming card. Meta's open weights are 816B total / 42B active. The smallest Unsloth 1-bit GGUF we ran (UD-IQ1_S) is 175.6 GB on disk and used ~168 GiB peak VRAM across 3x A100 80GB (r_fsd9qjmmna0). You need multi-GPU or multi-Studio class memory.
How fast is Muse Spark Open?
On 3x A100 80GB with Unsloth's 175.6 GB 1-bit file and llama.cpp 2851b3bc (build 11675) at 16K context, decode was 25.2 tok/s (chat-short), 25.5 (chat-long), 26.3 (concurrent-decode) and 22.8 (agent-trace), each a single run, not a median (r_fsd9qjmmna0). Prefill on the chat-long prompt (3,227 tokens) was 288.7 tok/s; first token took 11.2 s. We don't measure output quality.
How fast is Muse Spark Open on two Mac Studios?
Apple/MLX claim about 25 tok/s on two M5 Ultra Mac Studios and about 36 on four. Those are claims, unverified by us. We have no signed MLX Studio run yet. Submit a multi-Studio run via the request button and we'll measure it the same way.
Is 1-bit quality good enough?
We measure speed and memory, not quality. This page uses Unsloth's smallest 1-bit file (UD-IQ1_S, 175.6 GB) only. Higher quants are not in our set yet.
Related: day-one write-up · Nemotron 3.5 Bolt · Apple Silicon vs NVIDIA · best Mac for LLMs · best GPU 2026 · why is my local LLM slow
Local runs (4 workload results)
Runs from contributors' own machines via MLX, llama.cpp, vLLM, exllamav2, or ollama. Signed on the submitter's hardware.
3x A100 (80GB) + AMD EPYC 7742 64-Core Processor (128c) + 1008GB
| Workload | Backend | Quant | decode tok/s | prefill tok/s | TTFT | Run |
|---|---|---|---|---|---|---|
| chat-short | llama.cpp@0.7.0-dev (build 11675, commit 2851b3bcb) | IQ1_S | 25.20tok/s | 88.57tok/s | 2,732ms | r_fsd9qjmmna0 |
| chat-long | llama.cpp@0.7.0-dev (build 11675, commit 2851b3bcb) | IQ1_S | 25.50tok/s | 288.7tok/s | 11,179ms | r_fsd9qjmmna0 |
| concurrent-decode | llama.cpp@0.7.0-dev (build 11675, commit 2851b3bcb) | IQ1_S | 26.30tok/s | no data | no data | r_fsd9qjmmna0 |
| agent-trace | llama.cpp@0.7.0-dev (build 11675, commit 2851b3bcb) | IQ1_S | 22.79tok/s | 570.8tok/s | 3,896ms | r_fsd9qjmmna0 |
Muse-Spark-Open-816B-A42B-UD-00001-of-00005 on hardware
Request a Muse-Spark-Open-816B-A42B-UD-00001-of-00005 run
No measured run yet for these pairs. Vote for the one you want; the most-requested runs go first. See all requests.
We never sell or share your email. Privacy