Muse Spark 1.3 vs Muse Spark Open
| Benchmark | Muse Spark 1.3 | Muse Spark Open |
|---|---|---|
| Humanity's Last Exam | 47.5 | 43.9 |
| Terminal-Bench 4.0 | 16.7 | 8.1 |
| AutomationBench-AA | 56.8 | 49.7 |
| Arena Text · creative writing | 1,461 · max | — |
| MMMU-Pro (vision) | 82 | 80.9 |
Specs
| Spec | Muse Spark 1.3 | Muse Spark Open |
|---|---|---|
| Input, per 1M | $1.25 | $1.25 |
| Output, per 1M | $4.25 | $4.25 |
| Cost per task? | $1.37 | $0.84 |
| Speed, tok/s | 273 | 126 |
| Context | 1M | 1M |
| Weights | Closed | Open |
| Released | Sep 2026 | Oct 2026 |
| Memory (4-bit) | — | 493 GB |
| Licence | — | muse-model-license |
| Commercial use | — | Restricted |
| See images | ✓ | ✓ |
| Read PDFs | ✓ | Partial |
| Hear audio | — | ✓ |
| Watch video | ✓ | ✓ |
| Call tools | ✓ | ✓ |
| Force a tool call | — | Partial |
| Return strict JSON | ✓ | ✓ |
| Search the web | ✓ | ? |
| Run code | ? | ? |
| Use a computer | ? | ? |
What breaks when you switch
Reasoning off leaves output unchanged · Self-hosted / llama.cpp›
In llama.cpp, --reasoning off leaves the output unchanged; set reasoning_effort low to spend fewer thinking tokens.
Memory beyond the weights · Self-hosted›
Serving requires extra memory for KV cache, activations and runtime overhead; budget roughly 2 TB of VRAM for BF16.
Only auto tool_choice is supported›
tool_choice none, required or a named function returns HTTP 400; use auto and handle either a tool call or text.
Search filters return HTTP 400 · Messages API›
allowed_domains, blocked_domains and max_uses return HTTP 400 on Messages search grounding; omit them and use the supported web_search tool.
Manual budget does not set effort · Messages API›
Compatibility budget_tokens does not set reasoning effort; use output_config.effort for depth control.
No Chat Completions search grounding · Chat Completions›
Search grounding is unavailable in Chat Completions; use Responses or Messages.
Reasoning cannot be disabled›
Muse Wonder and Muse Spark do not support disabling reasoning with none; keep reasoning enabled.
Private reasoning still costs tokens›
Private reasoning consumes the output budget and billed output tokens; budget for it even though it is absent from the response body.