Two more model tracks rounded out the 2026 frontier. xAI shipped Grok 4.5, a closed model built for coding and agentic work, while Alibaba pushed its open-weight Qwen family forward with the natively multimodal Qwen 3.5 and the coding-focused Qwen 3.6. They sit at opposite ends of the open-versus-closed spectrum, which is exactly why both are worth tracking.

What happened

xAI Grok 4.5 was announced on 8 July 2026 as xAI’s most capable model, aimed at coding, agentic tasks, and knowledge work, with a 500,000-token context window. It is available through the API and the Grok apps; xAI does not release open weights for its frontier Grok models.

Alibaba’s Qwen went the open route. Qwen 3.5 (announced 15 February 2026) is a natively multimodal mixture-of-experts flagship, roughly 397 billion total parameters with about 17 billion active, spanning some 201 languages, with open weights. Qwen 3.6 (Qwen3.6-35B-A3B, 15 April 2026) is a smaller open-weight MoE (35 billion total, 3 billion active) focused on agentic coding, released under Apache 2.0. Note that Alibaba’s top tier, Qwen3.7-Max, went the other way as a proprietary, API-only model; the open weights cover the 3.5 and 3.6 families rather than the flagship.

Why it matters for builders

The split is the point. Grok 4.5 is a capable closed model you reach through an API, which suits teams that want performance without running infrastructure. Qwen 3.5 and 3.6 are open weights you can download, fine-tune, and self-host, which suits teams that need control, data residency, or offline deployment, and their mixture-of-experts design keeps inference cost well below what the total parameter counts suggest.

For a 2026 stack, Qwen’s open family plus DeepSeek-V4 and NVIDIA’s Nemotron 3 mean the open-weight tier is now genuinely frontier-adjacent, not a compromise. Choose by whether you need to own the model or just call it; the 2026 LLM landscape lays out the trade-offs.

Sources

Further reading