Meta dropped the weights for Muse Spark Open on October 15, and it's sitting at ~800 points on Hacker News. Most of the takes are either "Meta is back" or "that's not open source." Both are lazy. Let's look at what you actually get.
The spec sheet
- 816B total parameters, 42B active (mixture-of-experts)
- 1M-token context window
- Natively multimodal
- Aimed at coding and agentic work: long-horizon tasks, software engineering, multi-step reasoning
- Extra safety post-training meant to hold up when you run it bare, without system guardrails
- Custom Helion GPU kernels in SGLang: Meta claims 1.03-3.95x kernel-level speedups on NVIDIA and AMD, and a 3% end-to-end gain on AMD MI350
Meta calls it the most capable US open-weight model to date. Remember that "US" qualifier. It's doing work.
Benchmarks (vendor-reported, treat accordingly)
| Benchmark | Muse Spark Open | Best rival listed |
|---|---|---|
| MCP-Atlas (tool calling) | 89.2 | GLM 5.3: 84.2 |
| Terminal-Bench 2.1 | 76.4 | Kimi K3: 85.0 |
| DeepSWE 1.1 | 60 | Kimi K3: 68 |
| Intelligence Index | 39 | GLM 5.3: 45 |
Read that honestly: it wins at tool calling, and loses on raw coding and general intelligence to the Chinese frontier open models. For an agent stack, tool-call reliability matters a lot. For "write me a compiler," Kimi K3 still looks ahead.
HN commenters also pointed out that Artificial Analysis scores at the highest reasoning setting, which almost nobody uses in high-volume production. Tokens/sec and time-to-completion at realistic settings will tell you more than the headline index. Run your own evals.
The license: read it before you build on it
It's the Muse Model License, not Apache or MIT. The practical terms:
- Derivatives need a "Built with Muse Spark Open" attribution.
- If you offer model outputs as a service and make more than $20M a year in revenue, you need a separate written agreement with Meta.
- Internal use and powering your own proprietary product stay unrestricted.
For a startup, that's fine. For a hyperscaler or a big API reseller, that's the whole point of the clause. Call it "open weights", not "open source". The HN crowd is right on the label, even if the restriction is narrower than the outrage suggests.
What it takes to run
This is not a laptop model.
- NVFP4 checkpoint: ~700 GB VRAM minimum. Think four Blackwell Ultra or eight H200s.
- BF16: ~2 TB VRAM.
- 2-bit GGUF (Unsloth dynamic quant): fits in roughly 300 GB of combined RAM/VRAM.
- Meta's own pitch: two M5 Ultra Mac Studios with 512 GB unified memory each.
Serving with vLLM on 4 GPUs:
vllm serve meta-models/Muse-Spark-Open-816B-A42B-NVFP4 \
--tensor-parallel-size 4 --enable-expert-parallel
Or llama.cpp:
llama serve -hf ggml-org/Muse-Spark-Open-816B-A42B-GGUF:Q4_K_S
SGLang, TensorRT-LLM, MLX, and Transformers are also supported.
Don't want to own the hardware? It's already hosted on Together AI, Fireworks, Baseten, Modal, Groq, Nebius, and OpenRouter. Weights are on Hugging Face, SageMaker, Google Cloud Model Garden, and Microsoft Foundry. Pricing wasn't published in the sources I checked, so look at the endpoints directly.
My take
- This is an agent model first. The MCP-Atlas lead is the story. If you're building tool-heavy agents and want a non-Chinese open-weight option, this is now the one to eval.
- 42B active params is the lever. You pay VRAM for 816B but compute for 42B. Throughput per dollar on a properly sized cluster could be good. Test it, don't assume.
- The "US" framing says it all. Meta is no longer claiming the global lead. It's claiming the lead among American open models. That's a real concession about where open-weight frontier work is happening.
- Don't trust the table, including mine. Every number above is vendor-reported or secondhand. Build a 50-task eval from your own workload and run it against Muse Spark Open, GLM 5.3, and Kimi K3 before you migrate anything.
Should you care?
- Building agents with heavy tool use, and you have GPUs or a hosted endpoint: yes, eval it this week.
- Hobbyist with one consumer GPU: no. Even the 2-bit quant needs ~300 GB.
- Selling model outputs as a service at scale: talk to a lawyer first.
Sources: Meta research post, Hacker News thread, Developers Digest breakdown, PyTorch blog.
Top comments (0)