Chinese open-weight models have set the pace for downloadable AI for more than a year, and Meta now says it has an answer. On Thursday 15 October 2026 the company released Muse Spark Open, an open-weight model it claims “rivals, and in some cases outperforms” the best systems from Alibaba, Moonshot AI and Z.ai. Meta’s chief AI officer, Alexandr Wang, called it “the most capable US open-weight model to date” in a post on X, as AFP reported.
The claim is narrower than the headline suggests. Meta’s own comparison table shows the model leading on agentic tool use, but trailing the strongest Chinese open-weight models on coding and on the overall intelligence index. Artificial Analysis, which had published its own independent score by 16 October, puts it seventh among open-weight models, behind six Chinese open-weight models and ahead of every other American one.
This article looks at what Meta released, why Chinese open-weight models came to dominate the open market, and how the new model compares on benchmarks, licence terms, safety, hardware and price. We also cover what the contest means for UK businesses choosing a model to run themselves. For the full technical specification, see our earlier Muse Spark Open model guide. For Meta’s closed frontier models announced three days earlier, see our piece on Muse Wonder.
Table of contents
- What Meta Launched to Take On Chinese Open-Weight Models
- Why Chinese Open-Weight Models Lead the Open Market
- Meta’s Road From Llama to an Open Muse
- How Muse Spark Open Scores Against Chinese Open-Weight Models
- The US Response to Chinese Open-Weight Models
- How ‘Open’ Is Muse Spark Open?
- Safety Is Meta’s Sharpest Contrast With Chinese Open-Weight Models
- Running It: Two Mac Studios and an Apple Alliance
- What This Means for UK Businesses
- What to Watch Next in the Race With Chinese Open-Weight Models
- Frequently Asked Questions
- References
What Meta Launched to Take On Chinese Open-Weight Models
Muse Spark Open is the downloadable version of Muse Spark, the model family Meta Superintelligence Labs introduced in April. Unlike closed systems such as OpenAI’s ChatGPT or Anthropic’s Claude, an open-weight model lets developers download its parameters, or weights, and run or customise it on their own infrastructure.
The headline specifications
According to the Hugging Face model card, Muse Spark Open is a mixture-of-experts transformer with 816 billion total parameters, of which 42 billion are active for each token. Each layer holds 256 experts and routes every token to eight of them. The model has 62 layers, a 1-million-token context window and a 1.8-billion-parameter Perception Encoder for visual input.
It is natively multimodal: it accepts text, images, video and audio, and returns text. Meta says it was trained on data from more than 100 languages and is built for “long-horizon tasks, complex software engineering, and multi-step reasoning”. A reasoning-effort setting runs from low to xhigh, and xhigh is the default.
Where developers can get it
The weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry, according to Meta’s announcement as reported by The Hindu. Downloading from Hugging Face requires a manually approved access request. Hosted access went live on day one through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
Supported inference frameworks include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp, Unsloth and Apple’s MLX. That breadth matters, because the main appeal of Chinese open-weight models has been how easily developers can run them on whatever stack they already use.
The table below sets Meta’s model beside the three Chinese open-weight models it names as its targets, using the specifications MarkTechPost compiled from each vendor’s model card.
| Specification | Muse Spark Open | Kimi K3 | GLM 5.3 | Qwen3.8-2.4T-A95B |
|---|---|---|---|---|
| Developer | Meta (US) | Moonshot AI (China) | Z.ai (China) | Alibaba (China) |
| Total parameters | 816B | 2.8T | 753B | 2.4T |
| Active parameters | 42B | 104B | ~40B | 95B |
| Context window | 1M tokens | 1M tokens | 1M tokens | 262K native, up to ~1M |
| Input types | Text, image, video, audio | Text, image, video | Text | Text |
| Weights released | BF16, NVFP4 | MXFP4 | BF16, FP8 | BF16 |
Meta’s model is the smallest of the four by total size and the only one that accepts audio. That combination is central to its pitch: frontier-class agent work on hardware a single developer could own.
Why Chinese Open-Weight Models Lead the Open Market
Meta is entering a race it once led. Two years ago, its Llama family was the default choice for anyone building on downloadable models. Since then, the centre of gravity has moved to Chinese open-weight models.
Downloads have shifted east
Hugging Face’s State of Open Source report for spring 2026 found that Chinese open-weight models accounted for 41% of downloads on the platform between February 2025 and February 2026, against 36.5% for US models, as The New Stack reported. A joint MIT and Hugging Face study had already found China narrowly ahead, at about 17% against 15.8%, for the year to August 2025.
The ATOM Report, a separate study of the open language model ecosystem published in April 2026, measured cumulative downloads. It found that Chinese open-weight models overtook US ones in late July 2025 and reached 1.15 billion downloads by March 2026, against 723 million for US models. That is a gap of roughly 428 million downloads, up from 23 million in August 2025.
The chart shows Hugging Face’s download split for the year to February 2026. The remainder is the 100% total minus the Chinese and US shares (100 − 41 − 36.5 = 22.5).
Qwen became the default base model
Much of the lead held by Chinese open-weight models comes from one family. The ATOM Report credits Alibaba’s Qwen with almost a billion cumulative downloads by March 2026. According to Hugging Face data cited by The New Stack, models from Alibaba have spawned more than 100,000 derivatives. RunPod, which tracks deployments across its GPU cloud, found that Qwen had overtaken Llama as the most widely deployed self-hosted large language model.
Derivatives matter more than raw downloads. A team that fine-tunes a base model tends to stay with that family for its next project, so every Qwen derivative makes the next Qwen release easier to adopt. That network effect is a big part of why Chinese open-weight models are hard to dislodge. We covered one of the smaller members of that family in our Qwen3.8 27B guide.
From DeepSeek R1 to Kimi K3
The shift began with DeepSeek R1 in January 2025, which offered reasoning performance close to leading closed systems at a fraction of the price. Other labs followed, and the stream of Chinese open-weight models grew quickly. Baidu went from no Hugging Face releases in 2024 to more than 100 in 2025, according to Hugging Face, and ByteDance and Tencent each expanded their output by up to nine times.
By the autumn of 2026, the leading Chinese open-weight models included Moonshot AI’s Kimi K3, released in July, Z.ai’s GLM 5.3, Alibaba’s Qwen3.8 series and Xiaomi’s MiMo-V2.6-Pro. These are the models Meta is now measuring itself against.
Meta's Road From Llama to an Open Muse
Muse Spark Open is a reversal as much as a launch. Meta spent most of 2026 moving away from the open strategy that once defined its AI work.
Llama lost the lead
AFP noted that Meta’s open-source Llama models “lagged behind the rest of the industry” by the time the company changed course. Hugging Face’s own popularity data tells the same story: a year before the spring 2026 report, three of the five most-liked models on the platform were Llama models. By early 2026, DeepSeek-R1 held the top spot.
Spark went closed in April
When Meta launched Muse Spark in April, it was the company’s first major model without downloadable weights. Trending Topics described it as a break with the Llama strategy. Later point releases, Muse Spark 1.2 and 1.3, also shipped as closed, API-only models.
Meta did not abandon open weights entirely. In August it released Muse Glimmer, a 30-billion-parameter model under the permissive Apache 2.0 licence, aimed at laptops and personal computers. Glimmer was small, though, and was never meant to compete with the strongest Chinese open-weight models.
Wonder at the top, Spark Open below it
On Monday 12 October, Meta unveiled two closed frontier models, Wonder and Wonder Full, which it says rival the most advanced versions of ChatGPT and Claude. Three days later came Muse Spark Open. The pattern is now clear: Meta keeps its strongest models closed and releases a capable but deliberately lower tier as open weights to compete with Chinese open-weight models.
The gap between the tiers is visible in independent data. Artificial Analysis lists Meta’s closed Muse Spark 1.3 at 45 on its Intelligence Index, against 38.87 for the open version at the same xhigh reasoning setting. The Decoder also noted that Muse Spark Open “falls well short of Meta’s paid Muse Wonder Full models”.
How Muse Spark Open Scores Against Chinese Open-Weight Models
Meta published a comparison against seven other open-weight models of similar size. Three are Chinese: Qwen3.8-2.4T-A95B, GLM 5.3 and Kimi K3. The other four are Western: Thinking Machines’ Inkling, Nvidia’s Nemotron 3 Ultra, Reflection’s Beam and Mistral Large 4. All Chinese open-weight models in the table were run at their highest reasoning settings.
Where Meta’s model leads
The strongest results are on agentic tool use. Muse Spark Open scored 89.2 on MCP-Atlas, the top score in the table and ahead of Qwen3.8 (84.5), GLM 5.3 (84.2) and Kimi K3 (82.3). On JobBench it scored 61.8, narrowly ahead of GLM 5.3 at 61.4.
It also leads on two knowledge tests designed to reward accuracy over guessing. On HLE-Diamond without tools it scored 26.1, against 22.2 for Kimi K3. On AA Omniscience, which penalises confident wrong answers, it scored 25, against 20 for Kimi K3.
Where it trails: coding and the overall index
Coding is the weak spot, and coding is where Chinese open-weight models have built much of their reputation. On Terminal-Bench 2.1, Muse Spark Open scored 76.4, behind Kimi K3 (85.0), GLM 5.3 (83.9) and Qwen3.8 (82.0). On DeepSWE 1.1 it scored 60, against 68 for Kimi K3 and 67 for GLM 5.3. On SWE-Atlas-QnA it scored 58, against 66 for Kimi K3.
The table below shows selected rows from Meta’s own comparison. Higher is better in every row.
| Benchmark | Muse Spark Open | Qwen3.8 | GLM 5.3 | Kimi K3 |
|---|---|---|---|---|
| AA Intelligence Index | 39 | 40 | 45 | 44 |
| MCP-Atlas (tool use) | 89.2 | 84.5 | 84.2 | 82.3 |
| JobBench | 61.8 | 52.7 | 61.4 | 54.3 |
| Terminal-Bench 2.1 | 76.4 | 82.0 | 83.9 | 85.0 |
| DeepSWE 1.1 | 60 | 57 | 67 | 68 |
| GPQA Diamond (science) | 89.0 | 93.5 | 91.7 | 93.5 |
| HLE-Diamond (no tools) | 26.1 | 17.3 | 16.4 | 22.2 |
Counting wins and losses
Across the 20 capability rows in Meta’s table, excluding safety, we counted head-to-head results wherever both models report a score. Muse Spark Open beats Qwen3.8 on 8 rows and loses on 5. It beats GLM 5.3 on 4 rows and loses on 9. Against Kimi K3, which reports the most rows, it wins 8 and loses 12.
So “competes on a similar level” is fair against Qwen3.8 and arguable against the other two. “Outperforms” holds for tool use and some knowledge tests, but not across the board. The leading Chinese open-weight models still win more rows than they lose against Meta’s model.
Independent confirmation from Artificial Analysis
When the model launched, the figures came almost entirely from Meta, as Trending Topics pointed out. By 16 October, Artificial Analysis had published its own measurement: an Intelligence Index of 38.87, flagged as measured rather than estimated. That matches the 39 Meta cited.
The chart ranks the top open-weight models on Artificial Analysis’s index, read from its open-weights leaderboard on 16 October 2026. Each bar is scaled so that 50 index points fill the track.
Muse Spark Open is seventh. The six models above it are all Chinese open-weight models, and the gap to the leader is 7.4 points (46.3 − 38.9). The next American model, Reflection’s Beam, is 8.6 points behind Meta (38.9 − 30.3).
The US Response to Chinese Open-Weight Models
Meta is not the only American lab trying to close the gap. The Hindu noted that US companies including Meta and Nvidia have been racing to compete, partly to encourage wider adoption of American AI.
Beam, Nemotron and Inkling
Reflection launched Beam earlier in October, also pitched as a rival to Chinese open-weight models, as we reported in our Reflection Beam article. Nvidia offers Nemotron 3 Ultra, and Thinking Machines offers Inkling.
In Meta’s table, Muse Spark Open beat Inkling on all 18 rows they share and Nemotron 3 Ultra on all 11, often by more than 20 points. Against Beam it won six rows and lost two. On present evidence, Meta’s model is the strongest American open-weight release by a clear margin.
Mistral and the rest of the field
Outside both countries, France’s Mistral is the closest challenger. Mistral Large 4 scored 38 on the index in Meta’s table, a point below Muse Spark Open. Its weights are not yet public, so Artificial Analysis still lists it as proprietary, as we explained in our Mistral Large 4 ranking piece.
The broader picture is that a Western lab can now match the middle of the pack of Chinese open-weight models but not the top. That is progress from where Llama stood, but the leading Chinese open-weight models remain ahead.
How ‘Open’ Is Muse Spark Open?
AFP’s headline put “open” in quotation marks, and the licence explains why. Muse Spark Open ships under Meta’s own Muse Model License rather than a standard open-source licence.
What the licence allows
According to the licence text summarised by Trending Topics, developers may use, copy and distribute the model free of charge. Fine-tuning, derivative models and distillation are expressly permitted. Companies may use the model internally and build it into their own commercial products, and Meta claims no rights over the model’s outputs.
The strings attached
Two conditions stand out. Any product using the model or a derivative must prominently display “Built with Muse Spark Open”. And any company offering the model as a hosted AI service with more than $20 million in revenue over 12 months must notify Meta in advance and sign a separate agreement. That threshold does not apply to internal use, research or embedding the model in your own product.
The acceptable use policy bans weapons development, malware, operating critical infrastructure, fraud, disinformation and impersonation. The model card also states the model is not intended for anyone under 18.
How the licence compares with Chinese open-weight models
According to The New Stack, as cited by Trending Topics, the Muse licence closely resembles the one Moonshot AI uses for Kimi K3. Alibaba and Z.ai have historically released most Qwen and GLM models under permissive licences such as Apache 2.0 or MIT, with no revenue thresholds or naming rules, although the flagship versions in Meta’s comparison carry their own model licences.
| Licence term | Muse Model License | Apache 2.0 or MIT (many Qwen and GLM releases) |
|---|---|---|
| Free commercial use | Yes | Yes |
| Fine-tuning and distillation | Expressly permitted | Permitted |
| Mandatory attribution in product | “Built with Muse Spark Open” | Licence notice only |
| Revenue threshold for hosted services | $20M in 12 months | None |
| Use-policy restrictions | Yes, including weapons and critical infrastructure | None in the licence itself |
For most developers and start-ups, the terms are unlikely to bite. For large cloud and AI providers that want to resell the model, Meta remains the gatekeeper, which is a real difference from the permissively licensed Chinese open-weight models.
Openness beyond the licence
Artificial Analysis also scores openness, rating how much a lab discloses about its methods and data as well as its weights. Muse Spark Open scores 27.8 out of 100, with zero for both pre-training and post-training data transparency. The model card says only that training data was drawn from public sources, crawled from the internet, bought from third parties, derived from Meta’s products and services, and generated synthetically.
The chart compares that score with the leading Chinese open-weight models on the same leaderboard. Higher means more open, and the full track is 100 points.
On this measure, the American challenger is the least open of the group. The weights are downloadable, but the recipe is not.
Safety Is Meta's Sharpest Contrast With Chinese Open-Weight Models
Where Meta claims a clear lead is safety. The model card says the open version received “additional safety post-training to make it more robust against misuse as a standalone model without system guardrails”. That matters for open weights, because anyone who downloads the model can run it without the filters a hosted service would add.
Refusal rates against misuse
On StrongREJECT v2, an adaptive jailbreak test where lower is better, Meta reports a score of 0.5 for Muse Spark Open, against 65.6 for Qwen3.8, 62.9 for GLM 5.3 and 66.6 for Kimi K3. On cyber misuse, Meta says its model refused 99.2% of jailbreaks on an internal benchmark and 94.8% across three public ones, while Kimi K3 refused 75.5% and 70.2% in the same setup. Those are the widest gaps anywhere in Meta’s comparison with Chinese open-weight models.
For chemistry and biology, the card reports refusal of more than 98% of harmful requests on SecureBio’s BioTIER benchmark. On misalignment tests, it reports a 98.1% honesty rate on MASK.
The over-refusal trade-off
The same safety table shows the cost. On the FORTRESS benchmark, Muse Spark Open has the lowest adversarial risk score in the comparison, at 5.8. But its over-refusal score, which measures how often it declines harmless requests, is 17.4, against 12.8 for Qwen3.8, 3.2 for GLM 5.3 and 2.7 for Kimi K3.
Put simply, Meta’s model is far harder to misuse, but it is also more likely to refuse a legitimate request. Teams comparing it with Chinese open-weight models should test their own prompts, especially in security, chemistry or healthcare work. Note too that all these safety figures are Meta’s own. The model card warns that downstream fine-tuning “could degrade safety”.
Running It: Two Mac Studios and an Apple Alliance
The most unusual part of the launch was a hardware partnership. Meta says Muse Spark Open is “the first model in its class to run on your desktop”.
What the hardware has to hold
At 816 billion parameters, the model is smaller than Kimi K3 or Qwen3.8, the two largest Chinese open-weight models in Meta’s comparison, but it is still far too big for an ordinary workstation. Per Meta’s developer guide, as summarised by MarkTechPost, the full BF16 weights are about 1.6 TB and the NVFP4 version is about 530 GB. The model card recommends at least 700 GB of GPU memory for NVFP4 and 2 TB for BF16.
In a data centre, the NVFP4 build runs on four Nvidia Blackwell Ultra GPUs or eight H200s. On a desktop, it runs on two Mac Studios with M5 Ultra chips and 512 GB of unified memory each, linked over Thunderbolt 5 using Apple’s MLX framework. Two machines at 512 GB give 1,024 GB, enough for the 530 GB weights plus working memory. Unsloth’s 2-bit versions shrink the requirement to about 300 GB, with some loss of quality.
An unusual partnership
Mark Zuckerberg wrote on X that the model “runs great on new Mac Studios” thanks to work with Apple and its chief executive John Ternus, who succeeded Tim Cook in September. Ternus replied that it was “amazing to run a model of Muse Spark Open’s calibre and scale this efficiently, right from your desk”.
Wang told Axios the two companies had worked together for several months, crediting former Apple executive Daniel Gross with bringing them together. “We’re excited for mommy and daddy to be playing nice again,” he said. None of the Chinese open-weight models in MarkTechPost’s comparison has a disclosed desktop setup of this kind. As Mashable noted, Zuckerberg was criticising Apple on Joe Rogan’s podcast as recently as January 2025.
What it costs to use through an API
For teams that do not want to buy hardware, OpenRouter lists Muse Spark Open at $1.25 per million input tokens and $4.25 per million output tokens. Artificial Analysis calculates a blended price of $0.78 per million tokens and a median output speed of about 103 tokens per second.
The chart compares Artificial Analysis’s blended prices for the same group of models. Bars are scaled so that $2.50 per million tokens fills the track, and lower is cheaper.
Meta’s model is cheaper than Kimi K3 and close to GLM 5.3, but several Chinese open-weight models cost less than a quarter as much. For high-volume work, price may matter more than a few benchmark points. Our LLM API pricing guide covers how to compare these figures fairly.
What This Means for UK Businesses
For a UK organisation choosing a model to run itself, Meta’s release adds a credible American option to a shortlist that until now was dominated by Chinese open-weight models.
Provenance and procurement
Some organisations, particularly in the public sector, defence supply chains and regulated industries, prefer not to depend on Chinese-developed models, whatever their licence. Until now that often meant accepting a weaker model. Muse Spark Open narrows that gap, though the strongest Chinese open-weight models still score higher on coding and the overall index.
Self-hosting for data control
Running weights on your own hardware means prompts and documents never leave your environment. Wang told Axios that Meta has no visibility into developers’ local use of the model. That is the same benefit the Chinese open-weight models offer, and it is the reason many UK firms look at open weights at all. Our guide to private AI for UK businesses explains the options.
A practical evaluation checklist
Before switching from one of the Chinese open-weight models, or adopting either, we would test these points:
- Your own workload. Run the same 50 to 100 real tasks through each candidate. Meta’s lead on tool use and its deficit on coding may or may not matter for your work.
- Refusals. Check whether Muse Spark Open’s higher over-refusal rate blocks legitimate requests in your domain.
- Licence fit. Confirm you can display “Built with Muse Spark Open” and that you will not resell the model as a service above the $20 million threshold.
- Total cost. Compare API prices with the cost of 1 TB of GPU or unified memory, plus the staff time to run it.
- Exit route. Keep prompts and evaluations model-agnostic, so you can move between Meta and Chinese open-weight models as the rankings change.
If you are building an AI roadmap and want help choosing between open and closed models, our AI strategy service can run that evaluation with you. For a wider view of the market, see our guide to open-weight AI models in 2026.
What to Watch Next in the Race With Chinese Open-Weight Models
The contest is moving quickly, and several things could change the picture within weeks.
A detailed evaluation report
Meta says a fuller evaluation report will follow. Independent coding results, beyond the vendors’ own harnesses, will show whether the gap on Terminal-Bench and DeepSWE holds up.
The next wave of Chinese open-weight models
The leading Chinese open-weight models have been updated every few months. Any new Kimi, GLM or Qwen release could widen the gap again before Meta ships its next open version.
Mistral’s weights and US policy
Mistral has promised to publish the Large 4 weights, which would add a European contender at a similar level. In the US, policymakers and investors, including Y Combinator’s Garry Tan, as we covered in his call for US labs to distil frontier models, are pressing for stronger American open releases.
Frequently Asked Questions
What is Muse Spark Open?
It is Meta’s open-weight AI model, released on 15 October 2026. It has 816 billion total parameters, 42 billion active per token and a 1-million-token context window, and accepts text, image, video and audio input.
Is Muse Spark Open better than Chinese open-weight models?
Partly. It leads on agentic tool-use tests such as MCP-Atlas and JobBench, but trails Kimi K3, GLM 5.3 and Qwen3.8 on coding and on the Artificial Analysis Intelligence Index. It ranks seventh among open-weight models, behind six Chinese open-weight models.
Is Muse Spark Open open source?
Not in the strict sense. The weights are downloadable under Meta’s Muse Model License, which requires a “Built with Muse Spark Open” notice, sets a $20 million revenue threshold for hosted services, and includes use restrictions. Training data has not been published.
Can I run Muse Spark Open on my own computer?
Only on very large systems. Meta’s desktop setup uses two Mac Studios with M5 Ultra chips and 512 GB of memory each. In a data centre, it needs at least 700 GB of GPU memory for the NVFP4 version.
How much does Muse Spark Open cost to use?
The weights are free to download. Through OpenRouter, it costs $1.25 per million input tokens and $4.25 per million output tokens. That is cheaper than Kimi K3 but dearer than several other Chinese open-weight models.
Why are Chinese open-weight models so popular?
They combine strong benchmark results, low prices and, in many cases, permissive licences. Alibaba’s Qwen family alone has spawned more than 100,000 derivative models on Hugging Face.
References
Meta unveils new ‘open’ AI model to rival Chinese competition (AFP via The Star)
Muse Spark Open model card (Hugging Face)
Meta’s Muse Spark Open Is an Open-Weight Model With Strings Attached (Trending Topics)
Muse Spark Open is Meta’s answer to Qwen, GLM, and Kimi in the open-weight race (The Decoder)
Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context (MarkTechPost)
Meta launches new Muse Spark Open model, optimized to run on Mac Studio (9to5Mac)
Open-weight model leaderboard (Artificial Analysis)
China is winning the open source AI race (The New Stack)
The ATOM Report: Measuring the Open Language Model Ecosystem (arXiv)
Apple and Meta team up on Muse Spark Open for Mac Studio (Mashable)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
