muse – Teknotum

Meta releases open source Muse Glimmer, promises weights for Spark

Meta is making open source models central to its AI vision, says Zuckerberg. (Picture: Meta)

Meta claims to have squeezed a 30 billion parameter model onto a «consumer GPU» through clever 4-bit compression, making it weigh just 20GB, needing only 24 GB VRAM.

It not only runs agents, they say, but at a speed and responsiveness that feels natural without long thinking breaks.

At the same time, Mark Zuckerberg is out with a lengthy essay extolling the virtues of open models as central to their strategy, and presenting a «positive vision» for personal superintelligence as a tool for «individual empowerment.»

The selling point for Muse Glimmer is that it can run on a Pro- or Max-level MacBook Pro with sufficient memory, or on the GTX 4090/5090 GPUs from Nvidia on PCs.

These are high-end tools you likely won’t find in a corner store in Kampala, but run significantly below the cost of infrastructure-level Nvidia chips.

As an open model, Meta is only showing comparison benchmarks for Gemma 4-31B and Qwen 3.6-27B, which it seems to beat handily. It’s only barely showing on the LMArena leaderboards, however, at 97th for text and 77th for WebDev, below Gemma 4 and DeepSeek v3.2, but above Qwen 3.5.

You can fetch the model at Hugging Face under an Apache 2.0 license.

Read more: Meta’s presentation, launch post on X. Writeups on CNBC and TechCrunch. Discussions on Hacker News and r/LocalMMaMA.

Meta releases Muse Code in beta, competing on price and performance

Muse code competes primarily on price and general «good enough» quality. (Picture: Meta)

The new coding model from Meta comes close to Opus 5 in Claude Code while costing less than half as much to run, at $1.25 per million input tokens and $4.25 in output. It is available today in beta.

— Muse Code takes on complex software engineering tasks across large repositories: planning changes, writing code, and validating the results, Meta says.

It can coordinate subagents that are persistent and run in the background through each session, instead of respawning for individual tasks, reducing latency and feedback loops.

The agents also use advanced logging of every call, tool run, approval and edit, meaning they should run directly after restarts and crashes, picking up from the latest log entry like nothing happened.

The underlying model, Muse Spark 1.2, scores well on benchmarks for its price point. It clocks in just behind Opus 5’s 86.7% on Terminal-Bench with 82.9%, beating GPT-5.6 Terra and Gemini 3.6 Flash. On DeepSWE it scores 59.3%, ticking in behind GPT Terra and Opus 5 Max. On Artificial Analysis, a more general benchmark, it scores a little behind GPT-5.5.

— Muse Spark 1.2 was extensively trained on long-horizon coding tasks, including whole-repository generation, large end-to-end projects, and auto-research, Meta says on the model.

Muse Code is available with a «one-line install» from the terminal on macOS and Linux, and is available on dev.meta.ai on the web.

Read more: Meta’s launch page, X post by Zuckerberg, X post from Artificial Analysis. Writeups on CNBC, TechCrunch, and Reuters. Discussion on Hacker News and r/Singularity.

Meta kills the Instagram tagging feature on Muse Image after criticism

Letting anyone tag Instagram profiles for AI images had some obvious implications. (Picture: Meta)

The feature letting users take content from any public Instagram feed in AI mashups through newly released Muse Image promptly received backlash from concerned users, pressure groups and actors’ unions.

The idea was that you should be able to tag friends, but celebrities instantly started thinking about protecting their livelihoods and likeness — crucial to their success.

Even the National Center on Sexual Exploitation got on with the criticism, saying it was an obvious tool for sextortion and scammers, spotted by The Verge.

The actors union SAG-AFTRA quickly posted a guide to opting out, pointing to toggles deeply buried in the Instagram settings, while the Creative Artists Agency urged Meta to take a more «reasonable» approach and at least make the feature opt-in, instead of free-for-all.

After a tumultuous week online, Meta finally ceded to its critics, putting out this statement:

— Our intent was to provide a useful creative tool and to give people control over whether their public content could be referenced in this way. We’ve heard the feedback that this feature missed the mark, so it’s no longer available.

Read more: Updated Muse Image blog post, Deadline, Variety and The Verge.

Meta launches Muse Image generator, teases video model

A sample from Muse Image, Alexandr Wang as a plushie. (Picture: Meta)

Muse Image catapulted directly to the second place on Arena.ai’s leaderboard in the categories for Text-to-Image and Image Edit, ahead of Nano Banana 2 and just behind GPT-Image-2.

It’s the first image generator from the Superintelligence Lab and pairs with Muse Spark to reason, search the web, and plan before it generates an image.

Meta also says it knows you from your Facebook and Instagram feeds, giving it a layer of contextual awareness, and it lets you tag any Instagram account for inclusion in your images.

The latter is automatically enabled, meaning anyone on Meta.ai can tag your Instagram content in AI generations, unless you explicitly opt out.

Continue reading “Meta launches Muse Image generator, teases video model”