Right at midnight, quantized versions from several outfits went up, and when I downloaded unsloth's 4-bit version and tested it, I got a little over 11 tokens
Testing 512, 1024 and 2048 tokens ten times each gave 10~11.6 tokens, averaging about 11 tokens
On the off chance, I looked for an mtp version
Ollama had several versions up, so I downloaded one right away and tested it
Including 1024 and 2048, I confirmed it comes close to 30 tokens
So on the off chance I hooked it up to OpenCode and tested it
With one hooked up the speed drops slightly, but it still comes out fairly decent
They say mtp has gotten much better lately; for actual use 4-bit seems good and 8-bit seems more reliable, but if an nvfp4 version with mtp applied comes out, I would like to try that
