If you happen to have an nvidia RTX 4090, you can try my fork [1] to have a JEV compatible decisions endpoint with qwen-3.8-27b while simultaneously serving a fast chat endpoint (chat completions, responses, anthropic compatible), both sharing the same base weights and both with dynamic LoRA loading. This means you can essentially serve many fine tuned variants at the time on one consumer GPU. I still need to upload my decisions LoRA to huggingface so you don't have to train it yourself. I should probably also add support for this new openai decisions API format as well.
I'm getting about 3k tok/s on prefill so it heavily depends on the input size. I hooked it up into a coding agent for shell permission checks, estimated shell execution time (buckets), goal pre-screening, subagent routing etc. and I have RTTs between 100-200ms. It can take noticeably longer for huge contexts because of the prefill speed but it can be cached so subsequent calls will be much faster.
I tried smaller decision models like laya (also with custom finetunes) but the accuracy was not really good (for the things I tested). Also I don't have any VRAM left on this GPU so i had to decide whether to host laya or qwen-3.8-27b but not both at the same time. Running decision models on a CPU will also be noticeably slower so I went down this route to have both combined with shared base weights.
I have been looking for a good local setup for a while now. Qwen 3.8 27b is really good for a dense model of this size IMO. I already had an RTX 4090 and I forked ninfer [0] with the obsession to squeeze everything out of this card for this model. Results: 149 tok/s decode speed (aggregate with concurrency about ~270 tok/s) with prefill speeds faster than 2500 tok/s. And all of this with full 262k albeit quantized context. Fast prefill speed is really important when launching multiple clients such as opencode or pi at the same time and especially if they launch subagents. This is why I also implemented a caching tier so computed contexts can be faster loaded from RAM (or disk). Speeds feel almost like with official SOTA openai or anthropic models.
Im currently measuring a pareto front in J/tok in order to set power limits of this card without sacrificing too much performance. Since we are talking about full power draw of ~480W which is fine during the day (with solar panels) but during night when the sun doesn't shine (even with a battery) I'd like to limit this a little bit.
Sure, just consumer grade products: an ~8yr old 16 core AMD Threadripper CPU + 64 GB RAM, Samsung NVMe drive(s) and obviously an nvidia RTX 4090 (watercooled). Only the GPU is important though.
[1] https://github.com/tensorninja/ninfer-4090