Summary
On 15 October 2026, Meta released Muse Spark Open, an open-weight version of its Muse Spark model, which had been proprietary since April 2026. Meta describes it as the most capable open-weight model from the United States.
Built for coding and long-running agent tasks, it has 816 billion parameters, of which 42 billion are active for each token. On Terminal-Bench 2.1 it scored well below China's leading open-weight models: Kimi K3, GLM-5.3 and Qwen3.8.
The release shows that global AI competition is shifting from which country has the most powerful closed model to whose open weights become developers' default. That raises a choice for India between building on imported open weights and investing in sovereign models.
WHY IT MATTERS FOR UPSC & STATE PCS
Meta's release of Muse Spark Open is a direct US response to Chinese dominance in open-weight AI. It came alongside an Indian Express editorial arguing that leadership will go to whichever side combines affordability and trust with safeguards. For India, which is funding indigenous models under the IndiaAI Mission, the question of what to build on has become urgent.
Standard News
The Default Is the Prize and It Cannot Be Taken Back
Meta's release of Muse Spark Open shows how the AI race is changing. For two years, attention focused on which company had the most powerful closed model, available only through paid access. Now the contest is increasingly about open-weight models: systems whose trained parameters anyone can download, run and adapt.
Whoever supplies the model developers reach for first will shape the tools, standards and businesses built on top of it. China's labs, starting with DeepSeek-R1 in January 2025, have led this race. On the Terminal-Bench 2.1 coding benchmark, Kimi K3 scored 88.3% while Muse Spark Open scored 33.3% in independent tests.
The Case for Open Weights
Open weights lower costs dramatically. A startup or a public institution can run a capable model on its own hardware, adapt it to local languages or tasks and avoid paying per query to a foreign provider. They also allow uses that closed providers' terms may restrict, notably in cybersecurity, where defenders need AI that can analyse vulnerabilities.
And they reduce dependence on a few companies controlling access. For a country like India, this is how AI spreads beyond large firms.
The Case for Caution
Once weights are released, they cannot be recalled. A closed model can be updated, restricted or switched off; a downloaded model cannot. Powerful capabilities, including those that help attackers as well as defenders, become permanently available to anyone.
Policy has struggled with this. The United States issued a rule in January 2025 to control exports of certain advanced closed-weight models, then rescinded it in May 2025. Openly published software generally falls outside standard export controls, which shows how hard it is to govern weights once they are out.
There is also a dependence risk: building national systems on another country's models means inheriting its design choices.
Where This Leaves India
TAN's position is that India should build on open weights for broad adoption, while funding sovereign models in a few areas where dependence is unacceptable. Refusing imported open weights would deny Indian developers affordable tools that already exist.
But relying entirely on them would leave public services, Indian languages and sensitive domains dependent on models India cannot audit or update. The IndiaAI Mission's support for about 20 indigenous models, including Sarvam AI and BharatGen, should be concentrated where control matters most, not spread across a race to match frontier capability.
The capability India most needs is the ability to test, adapt and audit any model, imported or domestic, before it is deployed. That capability is the real basis of sovereignty in AI.
Quick Facts
Key numbers & takeaways — revise these first
-
Meta released Muse Spark Open on 15 October 2026; the original Muse Spark had been proprietary since April 2026.
-
Muse Spark Open has 816 billion parameters in a Mixture-of-Experts design, with 42 billion active per token and a context window of 1 million tokens.
-
It is released under Meta's Muse Model License, which allows open use of the weights with commercial restrictions.
-
On Terminal-Bench 2.1, Kimi K3 scored 88.3%, GLM-5.3 88.2% and Qwen3.8 Max 86.6%; independent tests put Muse Spark Open at 33.3%.
-
China's DeepSeek-R1, released in January 2025, made Chinese labs leading producers of open-weight models.
-
Under the Open Source Initiative's definition, models that release only weights, without training data and code, count as open-weight rather than open source.
-
The US AI Diffusion Rule, issued in January 2025 to control the export of some advanced closed-weight model parameters, was rescinded in May 2025.
-
The IndiaAI Mission supports about 20 indigenous foundation models, including Sarvam AI's 30-billion and 105-billion parameter models and the IIT Bombay-led BharatGen.
Connect the dots for your UPSC preparation.
Standard news covers the event. Log in to read our comprehensive analysis and uncover the hidden constitutional, structural, and ethical dimensions of this topic:
The full case for open weights: cost, local adaptation, cybersecurity defence and reduced dependence on a few providers, built at full strength
The full case for caution: why released weights cannot be recalled and why the US attempt at export controls was abandoned
TAN's position on India's strategy, sovereign models only where control matters and what would change that view
Why the ability to test and audit models may matter more to India than having a frontier model of its own
Included in this analysis
Join thousands of aspirants analyzing the news deeply.
Unlock Premium — Rs.699 AnnuallyDon't have an account? Sign up for free