Muse Spark Open vs Nemotron 3.5 Super VL

Muse Spark Open

Muse Spark Open runs on the owner's hardware in seven runtimes, with OpenAI-compatible servers in vLLM and SGLang, standard tool calls, a 1M-token context and Meta's own serving docs. Serving the 537 GB NVFP4 checkpoint takes four B300 GPUs, Meta's download is gated by manual review, and the licence adds attribution and a revenue threshold for model resellers.

Nemotron 3.5 Super VL

NVIDIA's free endpoint takes OpenAI-format requests with typed tool calls, JSON Schema output and an OpenAPI 3.1 file, and the weights run anywhere under a permissive licence. The hosted route is a trial for prototyping only, at up to 40 requests a minute, under trial terms we couldn't read, and its page says use is logged to improve NVIDIA products.

Muse Spark Open

1. Start vLLM or SGLang with --reasoning-parser muse --tool-call-parser muse (vLLM also --enable-auto-tool-choice), or tool calls come back as raw text
2. Set reasoning_effort to low or medium for short tasks. The default is xhigh, reasoning can't be turned off, and other values return an error
3. Stop only on <|eot|> and end of text. Stopping on <|eom|> cuts off tool calls and final answers
4. Stream long generations and leave max_tokens room for reasoning, or content can come back empty with finish_reason length
5. Don't call Meta Model API for this model. Use a self-hosted server or a hosted provider's model string, such as meta-models/muse-spark-open

Nemotron 3.5 Super VL

1. Set chat_template_kwargs to {"enable_thinking": false} for short tasks. Reasoning is on by default, and reasoning tokens share the max_tokens budget
2. Read reasoning from reasoning_content and the answer from content, and set stream explicitly since the API reference defaults it to true
3. Don't send confidential or personal data to the hosted trial. The model page asks callers not to and says use is logged
4. Back off and retry on throttling. Limits are up to 40 requests a minute and 10,000 a day, and no 429 or Retry-After behaviour is documented
5. For production, use a partner endpoint or serve the weights. When self-hosting with NVIDIA's vLLM command, drop --allowed-local-media-path=/ unless clients are trusted