When Meta pulled the plug on Llama in April 2026 in favor of the proprietary Muse Spark, the reaction was somewhere between panic and grief. A billion downloads, thousands of fine-tuned variants, an entire ecosystem built on Llama’s open weights — and overnight, no upgrade path.
Two months later, the story isn’t about what was lost — it’s about what rushed in to fill the void. China’s AI labs have seized the open-source mantle with a ferocity nobody predicted.
The Post-Llama Vacuum: What Meta Left Behind
Before the Muse Spark pivot, Meta was the loudest voice in the room for open-weight AI. Llama 2 and Llama 3 powered everything from weekend hobby projects to enterprise deployments at Fortune 500 companies. The models weren’t fully open-source (the Llama license had restrictive clauses around user scale), but they were downloadable, fine-tunable, and self-hostable — the de facto standard for anyone who wanted control over their AI stack.
Then Llama 4 landed in April 2025 with a thud. Independent benchmarks showed Maverick scoring 18 on the Artificial Analysis Intelligence Index — below models with half the compute budget. Behemoth was delayed indefinitely and entered vaporware status. The community turned hostile. Allegations of benchmark gaming surfaced.
Rather than iterate, Meta burned the bridge. In April 2026, Alexandr Wang’s Meta Superintelligence Labs shipped Muse Spark as a cloud-only, closed-weight model. No downloadable weights. No fine-tuning. No self-hosting. Andrew Ng called it “a significant loss for the developer community.” He wasn’t wrong — but he might have underestimated how fast the community would move on.
The Five Models That Now Run Open-Source AI
As of June 2026, five models dominate the open-weight LLM conversation. None of them come from American companies.
DeepSeek V4 Pro — The Workhorse
Parameters: 1.6 trillion total / 49 billion active (MoE)
Context: 1 million tokens
License: MIT
Pricing: $1.74 / $3.48 per million tokens (input/output)
DeepSeek V4 Pro is the direct heir to the Llama throne in spirit: massive scale, permissive MIT licensing, and a 1M-token context window that makes it the go-to choice for codebase-wide analysis and long-document processing. Its Flash variant runs on just 2× H100 GPUs at $0.14 per million input tokens, making it viable for budget self-hosting.
Qwen 3.7 Max — The Tool-Calling King
SWE-Bench Verified: 80.4% (tied with DeepSeek V4 Pro at 80.6%)
MCPMark: 37.0 — the best tool-calling score of any open model
License: Apache 2.0
Alibaba’s Qwen 3.7 Max debuted at #4 on the WebDev Arena leaderboard, ahead of Claude Opus 4.6, at half the price. Its standout feature is tool-calling reliability: in testing, Qwen 3.6 Plus produced correctly formatted tool calls on the first attempt 94% of the time. For AI agent frameworks — including Hermes Agent — this translates to fewer retry loops and more reliable autonomous execution.
Kimi K2.6 — The Swarm Architect
Parameters: ~1 trillion total / 32 billion active (MoE)
SWE-Bench Pro: 58.6%
License: Apache 2.0
Moonshot AI’s Kimi K2.6 introduced a genuinely novel capability: native swarm orchestration. It can internally spawn up to 300 sub-agents, managing task decomposition, delegation, and result synthesis within a single API call across up to 4,000 steps. For complex multi-file refactoring, it’s currently unmatched among open-weight models.
MiniMax M3 — The Newcomer (June 1, 2026)
Context: 1 million tokens
SWE-Bench Pro: 59%
Architecture: MiniMax Sparse Attention (MSA)
Promo Pricing: $0.30 / $1.20 per million tokens
Released just eight days ago, MiniMax M3 is the first open-weight model to combine three frontier capabilities in one package: top-tier coding, a 1M-token context window, and native multimodality. Its MSA architecture partitions the KV cache into blocks, cutting per-token compute at long context to roughly 1/20th the cost of the previous generation. At promo pricing, it’s absurdly cheap.
GLM 5.1 — The Dark Horse
Parameters: 744 billion total / ~80 billion active
SWE-Bench Pro: 58.4%
License: MIT
Zhipu AI’s GLM 5.1 is the strongest MIT-licensed model for long-horizon agentic coding. It matches Kimi K2.6’s SWE-Bench Pro score while being easier to self-host, making it the top pick for enterprises that need frontier performance with a permissive license and no legal ambiguity.
The Scoreboard: How They Stack Up
| Model | Total Params | Active Params | Context | SWE-Bench Pro | License | Input $/1M |
|---|---|---|---|---|---|---|
| DeepSeek V4 Pro | 1.6T | 49B | 1M | ~52% | MIT | $1.74 |
| Qwen 3.7 Max | Undisclosed | Undisclosed | 128K | 80.4%* | Apache 2.0 | $0.80 |
| Kimi K2.6 | ~1T | ~32B | 256K | 58.6% | Apache 2.0 | $0.60 |
| MiniMax M3 | Undisclosed | Undisclosed | 1M | 59% | Open-weight | $0.30 |
| GLM 5.1 | 744B | ~80B | 128K | 58.4% | MIT | $0.70 |
*SWE-Bench Verified (standard), not Pro
What This Shift Means for Developers
The implications of this landscape shakeup are profound:
1. The open-source LLM center of gravity has shifted to China. DeepSeek (Hangzhou), Alibaba/Qwen (Hangzhou), Moonshot/Kimi (Beijing), MiniMax (Shanghai), and Zhipu/GLM (Beijing) now collectively own the open-weight frontier. Not a single American company ships a competitive open-weight model as of June 2026.
2. Licensing has improved dramatically. Meta’s Llama license was always a compromise — commercial use with a 700M-user cap. The new guard ships under MIT (DeepSeek, GLM) or Apache 2.0 (Qwen, Kimi). For enterprises and startups, this eliminates legal uncertainty.
3. Pricing is in free fall. When Llama was the only game in town, API providers could charge a premium. Now, with five frontier models competing, input prices have crashed to as low as $0.14 per million tokens (DeepSeek V4 Flash). The days of $15/M-token proprietary pricing are numbered.
4. Specialization is replacing generality. Rather than one model to rule them all, we’re seeing specialization: Qwen for tool-calling, Kimi for swarm orchestration, DeepSeek for long context, GLM for MIT-licensed coding. Developers can now pick the right tool for each job.
5. Meta’s strategic error is now visible. By abandoning Llama, Meta didn’t just lose developer goodwill — it handed the open-source AI movement to Chinese competitors who are now setting the agenda. The billion Llama downloads Meta once celebrated are migrating to models that Meta doesn’t control, influence, or profit from.
What About the Llama Ecosystem?
The community hasn’t abandoned Llama overnight. Inference engines like llama.cpp remain essential infrastructure, and thousands of fine-tuned Llama variants still serve production traffic. But the writing is on the wall: without frontier-level updates from Meta, Llama models will increasingly lag behind.
For developers still running on Llama, the pragmatic path is clear: start evaluating alternatives now. Migrate non-critical workloads to DeepSeek or Qwen. Benchmark Kimi K2.6 for multi-agent use cases. Watch MiniMax M3 — if the promo pricing sticks, it could become the budget champion.
The Llama era was extraordinary while it lasted. But in AI, eras don’t last long. The open-source torch has been passed — and it’s burning brighter than ever.