Your local LLM predictions and hopes for May 2026
Community dreams of new Gemma, Qwen, DeepSeek, and custom chip models.
Deep Dive
A Reddit post by DeepOrangeSky asks the local LLM community to predict and rank models expected in May. The wishlist includes Gemma4 (124B), Qwen3.6 up to 397B, new GLM, Kimi, Nemotron, MiniMax, DeepSeek v4, Granite, Phi, and Taalas-style model-on-a-chip burners. It also asks about novel techniques like engram and surprising hardware players like AMD or Intel.
Key Points
- Predictions include models up to 400B+ parameters from Qwen and DeepSeek, plus new chip-on-model approaches like Taalas.
- Community wants improvements in neuromorphic techniques (engram) and surprise entries from AMD, Intel, or Samsung.
- The thread highlights strong demand for open-source local LLMs covering a wide range of sizes and specializations (coder, instruct).
Why It Matters
For professionals deploying AI locally, this signals a ramp-up in accessible, high-performance open models and novel hardware.