Open model gap: 80-160B models for large but slower memory devices
Recent open models have split into two groups. Some, such as 27B Qwen and 31B Gemma, fit machines with fast but smaller memory. Others, such as GLM 5.2, , Kimi 2.7, Mimo 2.5 Pro, and MiniMax M3, are very large models.
Devices with more than 96GB of memory, such as Apple machines, Ryzen AI 395 systems, DGX Spark, RTX 6000 Pro setups, four RTX 3090 cards, or 128GB DDR4/DDR5 memory, often have enough capacity but slower memory speed. These users must either run older 80-120B models that no longer feel close to the frontier, or fall back to smaller 26-35B models because is a bottleneck. The desired range is roughly 80-160B, especially a sparse model around 100B where only about 10B is active at a time.
That kind of model could fit AMD 9700 AI Pro systems, RTX 3090/5090 machines with 64GB VRAM, DGX Spark, AI395+ devices, and Apple machines.
Key points
- There is a gap for modern open models in the 80-160B range.
- Large but slower s are the main target hardware.
- Recent model releases are seen as either smaller fast-memory models or very large models.
- A sparse model around 100B with about 10B active is suggested as the practical shape.
- This could help reduce cloud if quality is good enough.