Gemma 4, Qwen 3.6 redefine local LLM performance

The arrival of Gemma 4 and Qwen 3.6 marks a shift where "frontier" performance is now consistently achievable on local developer workstations.

• –Qwen 3.6 35B uses a Mixture-of-Experts (MoE) architecture that enables 100+ tok/s on M5 Max, making it the superior choice for high-speed agentic loops.
• –Gemma 4 31B is a dense model prioritizing "intelligence-per-parameter," offering higher multimodal accuracy and creative reasoning at the cost of lower raw throughput.
• –Massive context windows (256K+) in both models allow for repository-level reasoning without cloud-based RAG overhead.
• –Apache 2.0 licensing for these weights ensures long-term viability for privacy-sensitive enterprise development.
• –Performance benchmarks show Qwen 3.6 dominating in coding (73.4% SWE-bench) while Gemma 4 leads in human-eval and multilingual tasks.