Key Takeaways
- Qwen (Alibaba) holds 69% of Hugging Face derivative model share versus Llama's 11%, down from Llama's 25% peak in November 2023
- Meta launches Muse Spark (April 8) as first closed-source model; abandons open-weight strategy after Llama 4 benchmark scandal and ecosystem loss to Chinese competitors
- Google releases Gemma 4 under Apache 2.0 (April 2) six days before Meta's closure, positioning as the 'enterprise-safe Western alternative'
- Chinese frontier open-weight models (GLM-5 at BenchLM 85, Qwen3.5 397B at 81) still outperform Gemma 4 on benchmarks despite compute constraints from US export controls
- China overtook US in cumulative Hugging Face downloads: 1.15B vs 723M, with Qwen's December 2025 single month exceeding combined downloads of next 8 model families
The Ecosystem Inversion: From Meta Leadership to Chinese Dominance
The structural dynamics of open-source AI have undergone a complete inversion in 18 months. Meta, which defined the open-weight era through Llama releases, has abandoned the strategy. Google, which historically kept models proprietary, is now the primary Western open-source provider. Chinese models, dismissed as derivative in 2023, now dominate the entire ecosystem.
The data is unambiguous. Qwen holds 69% of Hugging Face derivative model share versus Llama's 11%, down from Llama's 25% peak in November 2023. Qwen hit 700 million cumulative downloads by January 2026; its December 2025 single-month total exceeded the combined downloads of the next eight most popular model families. China overtook the US in cumulative Hugging Face downloads: 1.15 billion versus 723 million.
For ML engineers building on Llama, this signals immediate action required. The open-source ecosystem you relied on for fine-tuning, quantization tooling, and community support is no longer the dominant platform. Migration to Google's Gemma 4 Apache 2.0 or evaluation of Qwen is now a business decision, not a technical preference.
Hugging Face Open-Source Derivative Model Share (Feb 2026)
Chinese models (led by Qwen) now dominate the open-weight AI ecosystem that US companies created
Source: DEV Community / Humai Blog
Meta's Strategic Failure and Google's Positioned Response
Meta's pivot has three causes. First, the Llama 4 benchmark scandal: LeCun's admission that results were 'fudged a little bit' destroyed credibility. Second, the competitive failure: despite being open, Llama lost the derivative ecosystem to Qwen. Third, the $14.3B Scale AI acquisition brought Alexandr Wang as Chief AI Officer with a mandate to rebuild from scratch.
Muse Spark is the result—and it ranks 4th in composite benchmarks (score 52) versus Gemini 3.1 Pro and GPT-5.4 at 57. Meta went closed and got a less competitive model. The investment failed to produce either open-source ecosystem lock-in or proprietary model advantage.
Google's Gemma 4 Apache 2.0 release is strategically timed to exactly the moment Meta exits. The license upgrade (from restrictive custom license to Apache 2.0 with zero commercial restrictions) makes Gemma 4 the enterprise-safe alternative for on-premise deployments. The technical profile is strong: 89.2% AIME 2026, 92.4% MMLU, 256K context window, and extreme MoE sparsity (3.8B active of 26B parameters).
But the critical context is that Chinese frontier open-weight models still lead: GLM-5 scores 85 on BenchLM versus Qwen3.5 397B at 81, both above Gemma 4. Google is not the leader—it is the Western backup.
Frontier Model Competitive Landscape (April 2026)
Comparison of frontier models across key dimensions showing the new open-source vs closed-source fault lines
| Model | License | Open Weight | Key Strength | Composite Score |
|---|---|---|---|---|
| Gemini 3.1 Pro | Proprietary | No | Overall leader (tied) | 57 |
| GPT-5.4 | Proprietary | No | Coding (75.1 TermBench) | 57 |
| Claude Opus 4.6 | Proprietary | No | OSWorld verified (72.7) | 53 |
| Muse Spark | Proprietary | No | Health AI (42.8 HBench) | 52 |
| GLM-5 | Open | Yes | Top open-weight overall | 85 (BenchLM) |
| Qwen3.5 397B | Apache 2.0 | Yes | 69% HF derivatives | 81 (BenchLM) |
| Gemma 4 31B | Apache 2.0 | Yes | Enterprise-safe Western alt | Arena #3 open |
Source: Artificial Analysis / BenchLM / Arena AI / Lushbinary
Export Controls Accelerated Chinese Innovation
The geopolitical irony is profound. US export controls were designed to slow Chinese AI by restricting compute access. Instead, compute constraints forced Chinese labs into architectural innovations—extreme MoE sparsity, quantization efficiency, aggressive data curation—that made their models more deployable globally.
Chinese open-source AI now serves as the de facto global commons, while US labs retreat behind API walls. For enterprise buyers, the dilemma is real: the best open-weight models are Chinese-origin, but regulatory risk (supply chain controls, data sovereignty concerns, potential future restrictions) pushes toward Gemma 4 as the 'good enough Western alternative.'
This creates a two-tier open-source market: Chinese-led for performance, Google-led for governance. Your choice depends on whether you prioritize capability or regulatory certainty. For teams in regulated industries (finance, healthcare, government), the governance risk of Chinese models may outweigh the performance advantage. For teams optimizing for cost and capability, Qwen becomes the obvious choice.
What This Means for Practitioners
If you're building on Llama, evaluate migration paths immediately:
- For regulated industries: Default to Gemma 4 Apache 2.0. The license is permissive, the Western provenance reduces regulatory friction, and the 89.2% AIME score is production-viable for most applications
- For cost-optimized deployments: Evaluate Qwen3.5 397B (BenchLM 81) or Qwen's smaller variants. The performance advantage is significant, and the Chinese governance risk may be acceptable depending on your data classification
- On-premise and fine-tuning teams: Prepare for ecosystem tooling migration. Llama-specific infrastructure (quantization, inference optimization, training frameworks) will mature slower as developer attention shifts to Gemma 4 and Qwen
- API-strategy teams: Meta's closure reduces open-source competitive pressure on OpenAI and Anthropic. Expect API pricing to stabilize at higher levels, making on-premise deployments more economically attractive