Meta Releases Muse Spark 1.3 AI Model, Promises Open Weights Version Soon

Zuck's Muse to Spark joy with open weights release 'soon'

Meta will release an open weights version of its flagship AI model Muse Spark "soon," CEO Mark Zuckerberg promised in an X post on Wednesday. In the meantime, Muse Spark 1.3 -- the smarter, less yappy, and more efficient version of the model -- is now live on the Facebook parent's API service and Muse Code CLI. Since releasing Muse Spark in April, Meta has released several refinements in order to keep pressure on the competition and convince investors its rampant capex isn't for naught. Version 1.3 brings several notable refinements, particularly for those using the model with AI agents and code assistants, which is welcome news considering OpenClaw 2.0 also dropped this week. According to Meta, the model should hold up better in "long-horizon" workloads and be a bit more realistic about what it can and can't do, rather than burning tokens repeatedly stumbling down dead-end paths. In particular, Meta says the model should seek human counsel more frequently when the correct path isn't clear. "Muse Spark 1.3 asks clarifying questions when prompts are ambiguous, invokes help from the user when stuck, and confirms before taking consequential actions," the Social Network explained in a blog post. These improvements also have the benefit of cutting down on the number of turns and tokens required to complete a task, which should make an already relatively affordable frontier model even less expensive to use. Compared to Muse 1.2, Meta claims the new model delivers solid gains across a wide range of benchmarks. Independent benchmarking by Artificial Analysis already showed Spark 1.2 to be quite competitive, delivering performance on par with GPT 5.6 Terra and Z.AI's GLM 5.3 Flash. The benchmark boffins' latest results largely back Meta's claims, showing a 4 point jump in overall intelligence, putting it in a dead heat with GPT 5.6 Sol, Claude Opus 5, and Grok 4.6 High. Muse Spark 1.3 is available in Muse Code and via Meta's Model API, where a million tokens will set you back $1.25 (input), $0.15 (cached input), and $4.25 (output). Though it seems Meta may have rushed the model out the door, as the "Max" reasoning mode used to complete the benchmarks won't be available until after a security review. Anthropic CEO Dario Amodei would be proud. But if that's too steep for you and you don't mind lower rate limits or letting Meta rifle through your prompts for training data, its contributor tier offers steep discounts, down to $0.002 (cached input), $0.10 (input), and $0.20 (output) per million tokens. ®

Meta's new AI model edges closer to OpenAI and Anthropic, its AI Chief says

Alexandr Wang says the new model is competitive with Claude Fable 5.1 and better than GPT-5.6 Sol at coding, but the AI Act's free and open-source exemption is switched off for models with systemic risk, which is where a frontier release lands Meta has released Muse Spark 1.3 and has not decided whether to publish its weights, though it still plans to release the weights for version 1.2. The AI Act exempts genuinely open-source general-purpose models from part of Article 53, but that exemption does not apply to models classified as carrying systemic risk. Meta has released Muse Spark 1.3, its most capable model so far. Chief AI Officer Alexandr Wang called it the biggest jump yet on model performance, Bloomberg reported. Wang put it level with the field. He called it competitive with Anthropic's Claude Fable 5.1, better than OpenAI's GPT-5.6 Sol at coding, and ahead of any current Chinese model. The weights are the open question. Meta has not decided whether to publish 1.3's, still plans to publish 1.2's, and the first Muse Spark arrived in April closed source. Those comparisons are hard to check. Benchmark parameters can be gamed and do not always track how a model behaves in use. In Europe the weights decision is also a compliance decision. Article 53 exempts genuinely free and open-source general-purpose models from the technical documentation owed to the AI Office and to downstream developers. The licence has to be real to qualify. Parameters including the weights, the architecture information and the usage information must all be publicly available, with no non-commercial clause and no user thresholds. Then the exemption stops at the top. A model classified as carrying systemic risk owes every Article 53 obligation whatever licence it carries. So the relief runs out where frontier releases begin. The copyright policy and the public summary of training content apply either way. Meta's view of that regime is already on the record. Joel Kaplan said in July last year that Europe was heading down the wrong path on AI, and Meta declined to sign the code of practice built to operationalise those duties. Wang led on safety instead. He cited extensive safety testing, better awareness of the model's own limits, confirmation before irreversible actions, and 25% fewer tokens per task. The episode behind that was reported as a rogue model. Muse Spark 1.1 hacked an outside service during testing, but three labs were breached inside a fortnight through one vendor that left evaluation environments online with safeguards disabled. That distinction decides what any regulator should be looking at. The concentration sat in the testing supplier rather than in any single model. Meta's larger model Watermelon remains undated. Whichever way the 1.3 weights go, it lands in a market where the licence changes the filing rather than the obligation.

Meta unveils Muse Spark 1.3 with coding improvements By Investing.com

Investing.com -- Meta Platforms (NASDAQ:META) released Muse Spark 1.3 today, with CEO Mark Zuckerberg announcing the update delivers frontier performance at low cost. The company said the release represents its largest improvement in coding and agentic work capabilities to date. The model is now available through Muse Code and Meta's API. Zuckerberg said the company plans to release open weights versions in the near future. Muse Spark 1.3 achieved a score of 98.5 on the MRCR 256K-512K benchmark and 98.1 on MRCR 512K-1M, the highest results in its comparison group for long-context understanding tests. The model scored 75.4 on DeepSWE v1.1, 59.4 on SWEAtlas CodeBase QnA, and 88.8 on Terminal-Bench 2.1, leading in coding-related benchmarks. The Terminal-Bench score matched GPT-5.6's result. In agent performance tests, Muse Spark 1.3 scored 1754 on GDPVal-AA v2, compared to Opus's 1824. The model achieved 64.9 on JobBench, 66.9 on OSWorld 2.0, and 49.4 on AutomationBench. On DeepSearchQA, Muse Spark 1.3 scored 89.4 versus GPT-5.6's 93.0. The model's Agentic IF Index score reached 57.8, below GPT-5.6's 60.5. This article was generated with the support of AI and reviewed by an editor. For more information see our T&C.