Field note ·

Notes on Meta's Muse Spark 1.1: a real turnaround, and a few things I'm hoping to see clarified

Meta Superintelligence Labs shipped Muse Spark 1.1 this week, the public-preview follow-up to April’s Muse Spark 1.0, now available through the new Meta Model API. (Worth clearing up the name: the model is “Muse Spark,” the first in the “Muse” series, not to be confused with the old Spark AR platform or anyone else’s Spark.) I’ve been reading the reaction and trying it where I can, and I want to get my thoughts down, because this is a more interesting release than the skeptical first takes suggest, and Meta’s story over the last year is one of the more remarkable arcs in the space.

The turnaround itself deserves acknowledgment

Let me start with the thing I think gets undersold. A little over a year ago, Meta’s model story was in a rough spot after Llama 4. Since then they stood up MSL under a new org, brought in a lot of senior talent, and have now shipped two models in a single year that land in the frontier-adjacent conversation. Whatever you think of how the reorg happened, doing that in twelve months is hard, and the result isn’t vaporware. anthonypasq’s HN comment captured the surprise well: “The Alexander Wang acquisition, and all that it implies about how quickly and efficiently this team has built up their infrastructure and staffing, also seems legitimately impressive from a business development and engineering perspective.” It’s easy to be reflexively cynical here and miss that Meta genuinely re-entered the race.

The model itself is thoughtfully designed for where the field is going: natively multimodal, built for agentic work, with zero-shot use of new tools and MCP servers, multi-agent orchestration, a computer-use approach that writes scripts when automation is faster and clicks when direct interaction is simpler, and a 1M-token context. Meta claims a >10x compute-efficiency improvement over Llama 4 Maverick, which, if it holds up, is exactly the kind of unglamorous infrastructure progress that compounds. On capability, independent reads put 1.1 roughly in the Opus-4.6 / GLM-5.2 neighborhood for general agentic use, a notch below the very top on some coding metrics but genuinely in the mix. For a team rebuilding from where they were, that’s real progress.

The pricing signal is the pro-developer story

The part practitioners noticed most is cost. Meta is leaning into aggressive, low-cost pricing, and early hands-on impressions back it up. redox99: “very strong. It is actually much cheaper than Grok 4.5, especially cached reads.” Tiberium read it as “another indication that the BigLabs are feeling the GLM 5.2 heat.” I think that’s a good thing for everyone. A well-funded frontier lab competing hard on price keeps the whole market honest, and gives developers more room to build. Credit to Meta for showing up on the axis that actually affects day-to-day usage.

On going closed, I understand it, and I hope they revisit it

The emotional core of the reaction is that Meta, the lab that made open weights mainstream with Llama, shipped Muse Spark as a closed, API-only model. For a lot of people that landed as a loss. kilroy123 caught the mood: people building with “Meta’s local llama models have been the face of the open source AI scene for the past couple years… it didn’t seem likely that the scene was going to change very quickly, and then it did.” tpae framed it constructively, which is roughly where I sit: a major fumble that can still be fixed.

I want to be fair to Meta’s position. Shipping a frontier model open-weight is a genuinely fraught decision right now, competitive, safety, and regulatory pressures all push toward caution, and Meta’s blog does say they hope to open-source future versions if appropriate. So I read this as a defensible call under real constraints, not an abandonment of principle. But I’ll say plainly, because I think it’s the most useful feedback: Meta’s open-weight legacy is one of its best assets, and I hope they find a path back to it. Even a slightly-behind open release alongside the closed flagship would do enormous good for their standing with the exact developer community that roots for them.

The benchmark question, where I’d love more transparency

There’s a thread of skepticism about the benchmark numbers, partly a hangover from the Llama 4 Maverick episode, so trust is understandably still rebuilding. The careful version of the concern is worth surfacing: some commenters (meric_, solarkraft) noted that a Terminal-Bench result may have been run with raised resource limits while being reported under the standard benchmark name. As solarkraft put it, “resource limits are a part of the benchmark, changing them could change the benchmark,” adding, “I’m not trying to insinuate anything, this is just a thought…” I want to call this what it is: a community observation, not an audited finding.

My constructive ask is simple and achievable: publish the eval methodology alongside the scores, exact harness, resource limits, apples-to-apples config. phillipcarter’s “has there been any independent analysis?” is a fair thing to want, and it’s the kind of thing Meta can put to rest completely with a little more disclosure. Given they clearly have a real model this time, transparent methodology is pure upside, it converts skeptics instead of feeding them.

Where I’ve landed

  • The turnaround is real and underrated. Two frontier-adjacent models in a year, a strong efficiency claim, and a genuinely agentic design. Meta is back in the conversation, and that deserves acknowledgment.
  • The pricing is a gift to developers and healthy pressure on the whole market.
  • Going closed is understandable but, I hope, temporary. The open-weight legacy is Meta’s to reclaim, and I’d love to see them do it.
  • The benchmarks would benefit from more transparency. Publish the methodology and most of the skepticism evaporates.
  • My honest use: I’m putting 1.1 into real rotation for agentic and cost-sensitive tasks, including the 8B variants, and watching how it holds up beyond the benchmarks. The trajectory from 1.0 to 1.1 is encouraging.

It would have been easy to write Meta off a year ago, and Muse Spark is a good reminder not to. There are things I hope they clarify, the open-weight path and the benchmark methodology chief among them, but those are the asks you make of a team that’s back in the game, not one you’ve given up on. I’m rooting for the next one.

Read more