Recommended settings
The settings that most affect Muse Spark Open's output quality and throughput on a self-hosted server. Sampling, reasoning strength, and streaming are per-request, so tune them against a running server. Parallelism and context length are fixed by the serve command, so decide those before you launch.
Sampling
The recommended defaults are temperature 1.0, top_p 1.0, and do_sample true, which suit most use cases. The evaluation results in the Muse Spark Open model card(opens in new tab) were measured at these values with xhigh reasoning strength.
Reasoning strength
Muse Spark Open reasons before it answers. Reasoning strength sets how much it reasons, and it's the single biggest lever on both answer quality and token spend. A server started with the muse reasoning parser, such as SGLang or vLLM, returns the reasoning in a separate field from the answer.
On the request, reasoning_effort takes a level name. The open-weight chat template turns it into the numeric strength that the model reads from the system block:
reasoning_effort | Rendered strength | Use it for |
|---|---|---|
low | 8 | Classification, extraction, straightforward questions, and short tool loops |
medium | 32 | General agentic work |
high | 256 | Multi-step reasoning, hard debugging, and long tool loops |
xhigh | 512 (default) | Problems where depth matters more than latency |
The default is xhigh, so set reasoning_effort explicitly when you want shorter reasoning. Lower strength reduces generated reasoning tokens and latency.
Context length
Muse Spark Open carries a 1,048,576-token context window. Of its 62 decoder layers, 46 use sliding-window attention over the most recent 2,048 tokens and 16 use full attention. Only the full-attention layers keep a KV cache for the whole context, so KV memory grows with context length on those 16 layers and stays fixed on the rest. Set the server's maximum context length (--context-length in SGLang, --max-model-len in vLLM) to what your workload needs rather than the full window.
Parallelism
Muse Spark Open is a mixture-of-experts model, so expert weights dominate the footprint. Expert parallelism spreads those experts across ranks, and it's what makes the model fit on a given number of GPUs.
The serving recipes assume B300 GPUs.
| Checkpoint | Tensor parallel | Expert parallel | GPUs |
|---|---|---|---|
| BF16 | 8 | 8 | 8 × B300 |
| NVFP4 | 4 | 4 | 4 × B300 |
Parser flags:
- SGLang:
--reasoning-parser muse --tool-call-parser muse - vLLM:
--reasoning-parser muse --tool-call-parser muse --enable-auto-tool-choice - TensorRT-LLM:
--reasoning_parser musespark --tool_parser musespark
Stop tokens
Muse Spark Open ends a finished turn with <|eot|>. It ends a message that continues the turn with <|eom|>.
| Token | ID | Role |
|---|---|---|
<|begin_of_text|> | 200000 | bos_token_id |
<|end_of_text|> | 200001 | End of text, in eos_token_id |
<|eom|> | 200007 | Ends a message while the turn continues, such as reasoning, commentary, or a tool call followed by another call. Not in eos_token_id |
<|eot|> | 200008 | Ends the turn, in eos_token_id |
<|finetune_right_pad|> | 200018 | pad_token_id |
eos_token_id is [200001, 200008]. SGLang reads the same pair from the checkpoint's config.json and generation_config.json, so no explicit flag is needed there.
Stream long generations
Muse Spark Open produces long reasoning traces at higher strengths. Request streaming (stream: true) for reasoning workloads so a long generation doesn't hit a client or proxy timeout.
Next steps
Apply these settings in your runtime's serve command. See Run inference for SGLang, vLLM, TensorRT-LLM, and the other runtimes.