Broadly speaking, this ultimately pushes local inference towards a challenging world where you use SSD offload for weights as a matter of course; then smaller requests (or requests sharing the bulk of their context, e.g. subagent swarms) can be batched together and run quickly in aggregate, but running very large contexts will actually limit you to single-session inference and require swapping out even the KV cache itself to some external scratch SSD, further hurting your performance. Then feel free to add wide use of MTP in a probably futile effort to go back to tolerable tok/s numbers.
It is like a ping-pong game: the advantage flips back and forth between providers.
So yeah, it's the best local model I've seen. I am going to try the Qwopus 3.6 fine tune soon with the same spec and tickets and compare the output of both.
Not tried it yet but I've seen tests that suggest they've properly fixed the tool calling issues.
vLLM gives me ~7000+ tok/sec with Gemma 4's MoE model. Vs ~6000 tok/sec for Qwen 3.6 MoE.
But there’s also the quantization of DeepSeek v4 flash called dwarfstar