Meta Releases Muse Spark 1.3, Open-Weights Version Coming Soon

Meta has released Muse Spark 1.3, the latest version of its flagship AI model Muse Spark. Chief Executive Officer (CEO) Mark Zuckerberg said in a post on X that an open-weights version will also be offered soon. According to reporting by Tobias Mann of The Register, the remarks were made on September 2.

What is available at present is Muse Spark 1.3 provided via the API service and Muse Code CLI. Since first unveiling Muse Spark in April, Meta has made continuous improvements against a backdrop of competition and accountability to investors. This update is heavily focused on use in AI agents and code assistance. The fact that its release coincided with the arrival of OpenClaw 2.0 also illustrates the intensifying competition in the agent space.

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.

The above is a quote from Zuckerberg’s post as reported by The Register. The specific timing and licensing terms for the open-weights version have not been disclosed at this time.

Launch of Muse Spark 1.3

Muse Spark 1.3 is immediately available via Meta’s API service and Muse Code CLI. Since the initial release in April, Meta has rolled out several improved versions. Version 1.3 is positioned as an update centered on efficiency and dialogue control.

In its official blog, the company described the aim of 1.3 as improving resilience for long-running tasks. While previous behavior that consumed tokens on dead-end paths had been seen as an issue, the new version is said to improve behavior in uncertain situations. Maintaining both API and CLI as delivery channels for developers can be seen as a decision that prioritizes integration into existing workflows.

The stability of the delivery format is a factor that affects the developer experience. In recent platform updates, cases have been reported where environment-dependent bugs disrupted work, as in Microsoft Update Leaves Arm Versions of Outlook and Teams Unable to Launch. Continuing to offer access via CLI helps ensure reproducibility in local environments and compatibility with automation.

Enhanced Adaptation for Long-Running Tasks

and Improved Efficiency

The improvement Meta emphasizes is more realistic behavior on long-running tasks. The model has been designed to better recognize the limits of its own capabilities and, rather than continuing to make easy guesses in response to unclear instructions, to seek confirmation from the user.

Specifically, three behaviors are cited: asking clarifying questions when prompts are ambiguous, seeking the user’s advice when stuck, and confirming before taking consequential actions. According to reporting by The Register, Meta explained in a blog post as follows:

Muse Spark 1.3 asks clarifying questions when prompts are ambiguous, invokes help from the user when stuck, and confirms before taking consequential actions

These changes also lead to a reduction in the number of turns and tokens required to complete tasks. Meta says that Muse Spark, which has already been positioned as a relatively low-cost frontier model, will be even more cost-effective with version 1.3. In scenarios where an agent autonomously executes multiple steps, suppressing unnecessary attempts directly translates into cost savings. This is also a factor in the reliability when developers delegate long-running work to agents.

Benchmarks Show Parity with Rival Models

In terms of performance, improvements across a wide range of benchmarks are claimed in the transition from Muse Spark 1.2 to 1.3. As a third-party evaluation, the independent benchmark by Artificial Analysis is cited. In the organization’s previous results, Muse Spark 1.2 showed performance comparable to GPT 5.6 Terra and Z.AI’s GLM 5.3 Flash.

In the latest results, Muse Spark 1.3 recorded a 4-point increase on a comprehensive intelligence metric. According to The Register, this puts it roughly on par with GPT 5.6 Sol, Claude Opus 5, and Grok 4.6 High. A competitive comparison chart presented by Meta itself also supports this claim.

Benchmark scores do not determine a model’s value on a single metric alone, but they serve as material for gauging competitiveness at the frontier. Especially for agent use cases, not only reasoning accuracy but also long-context retention and handling of ambiguous instructions are important. How much the 4-point improvement in this update will be felt in such real-world use is expected to become clear as verification by developers progresses. Intensifying benchmark competition is also prominent on the hardware side, as seen in AMD Medusa Point Sets x86 Speed Record on Geekbench, with growing attention on performance evaluation for both models and the infrastructure that runs them.

Open-Weights Release Preview and Aims

The provision of an open-weights version, which Zuckerberg described as coming “soon,” was the most eye-catching element of this announcement. Meta has previously promoted open releases with its Llama series and strengthened collaboration with researchers and the developer community. Adopting a similar approach for Muse Spark can be interpreted as an intention to expand the ecosystem.

Open-weighting is an approach that publishes model weights, allowing users to run the model in their own environment and fine-tune it. Compared with API-based use, it broadens options for organizations that want to avoid sending data externally or that wish to adapt the model to specific domains on their own. At the same time, it brings challenges such as preventing misuse after release and license management. At a stage where Meta has not yet shown specific terms, the focus will be on the scope of the release and whether commercial use will be permitted.

From an investor perspective, there is constant scrutiny over whether Meta’s huge capital investments are translating into results. The continuous improvement of Muse Spark and the preview of an open-weights version can be seen as aiming to make R&D results visible in both products and the community. With competitors successively releasing new versions, it also strongly signals an effort to lock in developers.

Pricing Structure and Usage Options

On pricing, Muse Spark 1.3 offers a tiered pricing structure in addition to an effective cost reduction through improved efficiency. Meta provides multiple tiers depending on usage. For the contributor tier, rates are shown as $0.002 for cached input, $0.10 for input, and $0.20 for output per million tokens.

This tier comes with lower rate limits and the condition that prompts may be used as training data. While this is more acceptable for cost-conscious individual developers and research use, enterprise use handling sensitive data may require a higher tier with looser limits or a private offering. The approach of clearly separating pricing and data-use terms can be seen as intended to make it easier for users to choose according to their needs.

Improved token efficiency affects not only the unit price but also total cost. Reducing the number of turns also leads to fewer API calls and shorter wait times. In workflows where an agent makes multiple tool calls, curbing waste per call is effective for both overall stability and cost.

Developer Offerings and Future Focus

Muse Spark 1.3 is offered via two channels: the API service and Muse Code CLI. The API is suited for integration into existing applications and agent infrastructure, while the CLI is suited for local coding assistance and integration with automation scripts. That Meta updated both simultaneously indicates it anticipates use in both cloud and local environments.

The improvements for agent use come at a time overlapping with the release of OpenClaw 2.0, making comparisons more likely. Resilience for long-running tasks and confirmation behavior in uncertain situations affect the balance between agent autonomy and safety. Meta’s explanation that it has “increased the frequency of seeking human advice” reflects a design philosophy of curbing autonomy if necessary to prevent going too far down the wrong path.

The focus going forward will shift to the specifics of the open-weights release. Once details such as the scope of the weight release, license, required compute resources, and fine-tuning procedures are revealed, companies and research institutions will be able to concretize adoption plans. Based on operational results and benchmark outcomes for the API version, the question will be how much reproducibility the open version offers. Improvements on the desktop environment side are also progressing in parallel, and as seen in COSMIC Epoch 1.2 Fixes Intel Graphics Flickering, greater stability of developers’ overall working environments underpins the use of AI tools.

Editorial Opinion

In the short term, we expect the efficiency improvements and enhanced confirmation behavior in Muse Spark 1.3 to drive adoption on the front lines of agent development. Reducing tokens and turns directly lowers costs, and inserting confirmations for ambiguous instructions helps curb erroneous actions. Combined with updates to competing agent platforms such as OpenClaw 2.0, best practices for long-running agent execution could converge rapidly within 3–6 months. In the long term, we assess that the success or failure of the open-weights version will determine the evaluation of Meta’s overall AI strategy. If the weight release materializes, adoption will expand among organizations seeking domain-specific adaptation and on-premises operation. On the other hand, how much of the post-release operation and safety measures are entrusted to the community will affect the sustainability of the ecosystem over 1–3 years. Whether a two-tier structure of API and open offerings takes hold will depend on licensing and support. As a question from the editorial team, we believe the transparency of the model’s decision to “seek advice” is at issue. At what threshold confirmations are inserted and whether users can control that behavior relates to trust in real-world operation. Too many confirmations impair autonomy, while too few allow erroneous actions to proceed.

References

• “Zuck’s Muse to Spark joy with open weights release ‘soon’”, by Tobias Mann — The Register, 2026-09-02T22:31:10.000Z (ARR)
• Source URL: https://www.theregister.com/ai-and-ml/2026/09/02/zucks-muse-to-spark-joy-with-open-weights-release-soon/5294093