I know it’s an hybrid, but it’s the first non codex/claude model I’ve found that can complete a very complex engineering process from a design, end-to-end, without spinning into oblivion or crashing under lmstudio Bionic.
Even the open source gpt-20b didn’t survive.
Yesterday I had it running nonstop for 6 hours on my 5060Ti 16GB and it was flawless. Finished the task perfectly. My design spec is 26k, and it chewed through that in 12 seconds and never looked back. It iterated through 8 stages and 64 sub-stages like it was nothing. May not seem like a lot to many with more powerful hardware, but on limited capacity this is the best I’ve found. Would like to hear your thoughts.
Hey there, creator of the original model merge here. Just saw this comment and wanted to say thank you! Feedback from the community like this is very valuable, because it serves as a benchmark on real-world tasks with different hardware, OS, agent harness, drivers, etc. which it turn helps me optimize my merge "recipes" for future releases. If you're curious, I’ve got a few other hybrid merges that offer similar agentic coding capabilities you can try out: A 9B merge of MiMo-V2.6-Distill-Qwen, Ornith-1.5 and Qwopus3.5-Coder: https://huggingface.co/collections/pragmaticcs/triumvirate-qwopus-mimo-ornith-9b-coder A 35B-A3B Merge of Qwopus3.6-Coder, KAT-Coder-V2.5-Dev, Qwen-AgentWorld and Ornith-1.5: https://huggingface.co/collections/pragmaticcs/qwen-35b-a3b-signoffour-coder
And a huge shoutout to @mradermacher for putting together static and imatrix quants of this model!!!
