Muse Spark 1.1 Benchmarks: How Good Is Meta's New Model?
Meta has spent the past few years being better known for open Llama models than for competing at the very top of the AI market. Muse Spark pushes Meta back into that high-end competition.
Muse Spark 1.1 is a proprietary model available through an API. It can work with text and images, call tools, reason at different effort levels, and accept up to one million input tokens. Its strongest independent results are in coding and scientific reasoning.
It is not the best model on every benchmark. It is also not especially quick to start answering. What makes it interesting is the combination of strong performance, a very large context window, and inexpensive cached input.
The short version:
- Use Muse Spark 1.1 for coding, difficult analysis, and repeated work with large documents or repositories.
- Consider another model for instant chat, the hardest autonomous agent tasks, or work where only the very highest benchmark performance will do.
- Treat Meta's launch results as first-party evidence. Independent testing shows real progress, but also some important limits.
What changed in Muse Spark 1.1?
Muse Spark 1.1 arrived three months after Meta introduced the original Muse Spark. Artificial Analysis independently tested both versions and gave 1.1 a score of 51 on its Intelligence Index, up from 43 for the first release.
That places it near GPT-5.4 at xhigh reasoning, GPT-5.6 Luna at max, and GLM-5.2 at max in the same testing. Those labels refer to each model's highest reasoning setting. Muse Spark remains behind current leaders such as Claude Fable 5, GPT-5.6 Sol, and Claude Opus 4.8.
The improvement was not evenly spread across every kind of task. The clearest gains were in scientific reasoning, coding, and knowledge work. Muse Spark 1.1 also expanded its context window from 262,000 tokens to one million.
Unlike Meta's Llama releases, Muse Spark is not open weights. You use the hosted model rather than downloading and running it yourself.
The independent benchmark picture
Artificial Analysis tested Muse Spark 1.1 at its highest reasoning setting. These results show what the model can do with a large reasoning budget, not what every ordinary conversation will achieve.
| Independent result | Muse Spark 1.1 | What it suggests |
|---|---|---|
| Intelligence Index | 51 | Competitive with several upper-tier models, but below the leaders |
| Coding Index | 71 | Coding is one of its strongest areas |
| SciCode | 58% | Third-highest result in the tested comparison |
| Humanity's Last Exam | 45% | Strong performance on unusually difficult questions |
| GDPval-AA v2 | 1376 Elo | Better at professional tasks, but still behind the frontier |
The SciCode result stands out. At 58%, Muse Spark 1.1 trailed only Claude Fable 5 and Gemini 3.1 Pro Preview in Artificial Analysis's testing. Its 45% on Humanity's Last Exam was also within one point of Claude Opus 4.8's 46%.
Those numbers make Muse Spark a serious option for difficult coding and reasoning. They do not make it the obvious choice for every agent workflow. Its GDPval-AA v2 score rose from 1144 to 1376 Elo, a rating system where higher means better performance in head-to-head comparisons, but it still lagged the strongest models. Meta's own evaluation report also shows that it trails leading competitors on some long-running coding and agent tasks.
A model can be very good at solving a difficult programming problem without being equally good at managing an entire project through hundreds of steps, and the results suggest Muse Spark sits on the first side of that line.
Meta ran its own Muse Spark results at xhigh reasoning and notes that its setup may not have been tuned to every competing model's strengths. Its figures are useful, but they should not be read as a neutral guarantee for every workload.
Fast output, slow start
Artificial Analysis measured about 114 output tokens per second through Meta's API. Once Muse Spark begins answering, it produces text quickly.
The wait before that answer was much longer: roughly 21 seconds to the first answer token in the same testing. That can feel slow in a normal chat, even if the response arrives rapidly afterward.
Muse Spark therefore makes more sense for work where the answer itself matters more than an instant reaction. Reviewing a large repository, comparing several documents, or solving a hard technical problem can justify the wait. A customer-facing chat widget or quick factual assistant probably cannot.
Lower hallucination does not mean higher accuracy
One of the most interesting changes in 1.1 is how it handles questions it may not know.
On its difficult AA-Omniscience test suite, Artificial Analysis measured a drop in hallucination rate from 73% to 38%. That benchmark-specific figure is not a general hallucination rate for every conversation. The underlying result is also more specific than a broad accuracy improvement: the model attempted fewer questions, while its answer accuracy declined slightly from 45% to 41%.
In plain English, Muse Spark 1.1 became much more willing to admit uncertainty.
A careful refusal is often better than a confident invention, especially in research or professional work. But it would be misleading to say the model became dramatically more knowledgeable. It became more cautious about when to answer.
The practical value of a 1M-token context window
A one-million-token window can hold a large codebase, many reports, or an extended project history. Muse Spark also supports prompt caching on NanoGPT, which lowers the price when the same input is reused.
Current NanoGPT pricing is:
| Usage | Price per million tokens |
|---|---|
| Input | $1.25 |
| Cached input | $0.15 |
| Output | $4.25 |
The cached input price is especially useful for repeated work. If you load a large repository once and ask many follow-up questions about it, later requests can reuse the same material at a much lower input price. For a one-off prompt, caching offers little benefit.
Large capacity is not perfect memory. Important details can still be missed when they are buried among hundreds of thousands of tokens. It helps to tell the model which files or sections matter, ask it to summarize the evidence first, and request references to filenames or headings in the final answer.
Where Muse Spark 1.1 makes sense
Muse Spark is worth trying for:
- Debugging a difficult issue across several files
- Reviewing a large repository before planning a change
- Comparing long reports, policies, or research material
- Analyzing screenshots alongside written instructions
- Tool-using workflows with a clear goal and sensible limits
- Background analysis where a slower first response is acceptable
It is less compelling for:
- Quick chat where every second matters
- Short rewrites, summaries, or extraction that a cheaper model already handles
- Fully autonomous work that may run for a long time without supervision
- Tasks where you need the strongest model available regardless of price
The best fit is substantial but bounded work: give it a large body of relevant material and a clear problem, then review the result before moving to the next step.
Choosing a reasoning effort
Muse Spark 1.1 supports minimal, low, medium, high, and xhigh reasoning on NanoGPT. Reasoning cannot be completely disabled; even the minimal setting uses some reasoning budget.
Use minimal or low for straightforward summaries, extraction, and drafting. Start with medium for coding and ordinary analysis. Move to high or xhigh when the task is genuinely difficult or the lower setting misses constraints and interactions.
The published Artificial Analysis score of 51 used xhigh. Lower settings should be faster and may use fewer tokens, but they should not be expected to match the headline benchmark result.
More reasoning is not automatically better. If a task is simple, a larger reasoning budget can add waiting and output without improving the answer. Test the same real prompt at two effort levels before choosing a default for repeated work.
The bottom line
Muse Spark 1.1 is a meaningful return to high-end model development for Meta. Independent testing confirms a large improvement over the first Muse Spark, with particularly strong coding and scientific reasoning results.
It will not beat the frontier leaders on raw benchmarks, but it gets close enough for many serious tasks while offering a one-million-token context window, cheap cached input, and fast output once generation begins.
The main tradeoffs are a slow first response, weaker results on some long-running agent tasks, and benchmark performance that depends on using a high reasoning setting. If those limits fit your workflow, Muse Spark 1.1 is a practical model to test rather than just another launch to watch.
You can try Muse Spark 1.1 on NanoGPT. API users can select meta/muse-spark-1.1 through NanoGPT's OpenAI-compatible API.
