vblanco

Built in codemode is rather nice. Its the feature i like the most from OMP which is Pi + lots of plugins.

380.000 concurrent sandboxes on 160 Epyc based server nodes. Crazy stuff

Completely misread the github and understood this as a agent inside unreal as a plugin...

In my projects i use unreal and me and my team have tested out multiple things. We have settled on no MCP complications whatsoever other than simply exposing the python scripting from the editor + a export step that can write the blueprints and asset data into plaintext so that the bot can grep them. This has given by far the strongest results, and we now have no issues having the agents edit game code and do operations.

We found this massively outperforms any kind of agent like this and the official unreal MCP systems. Its similar to the Blender MCP which also just exposes scripting + very minimal api to claude code/others.

The models they are using are Nemotron3 series from last year. For reference at that time, the state of the art was Sonnet 4.6 and GPT was essentially unused for coding. Currently, you can run Qwen 27B in your gaming PC and it will absolutely trounce every model they show in their report.

This is a field that changes significantly every few months, so using 1 year old models essentially invalidates the entire report, as they are 2 model generations behind and the newer generation models are heavily RL-d around their harness. For example they put a lot of emphasis on context management but that kind of context management isnt done anymore because many models now have 250k or 1M context and inferencing the models heavily rewards cache hits so you should never touch the context until you decide to compact. They comment on plan mode, which is something that Claude decided to remove from their client because they dont consider it needed at all anymore.

This is done on Nemotron models + mistral, so its not very relevant to the current frontier of cheap chinese models + big models from Claude/GPT. Big miss not having qwen or deepseek in this research.

Insane scores for a model of this size. But it does seem to be a rather insane over-thinker with the biggest token use of any model, which combined with 256k context size it means it wont do much before filling it context

Its free tokens so i left it running for fun as a experiment

completely vibes based, but ive been using it to port Mindustry game from Java to C# with agents, and its been working for 50 hours (its 15-20 tks so super slow inference). Its done a fantastic work and its almost finished now. Better results than deepseek flash and gpt luna by a mile on this kind of long term work. Less good than gpt sol or opus. We dont know the param count but my guess is 200-300 range.

just the default one in the link for image-to-video

Im running this on my 4070ti super (16 gb vram), and it takes 10 minutes for a 10-seconds 480p video. but the results are spectacular.