Has anyone actually tried this on real shop floor drawings?
Hey folks, stumbled across this model and it looks pretty sick — process planning straight from a 2D drawing is exactly the kind of thing I've been wishing existed for ages.
A few quick questions for anyone who's actually played with it:
- How does it handle handwritten notes on old PDFs? Like, if a machinist scribbled "tap M6 here" in the margin, does it pick that up or does it need super clean drawings?
- Any idea what kind of GPU I'm looking at to run this locally? My workstation's got a 3090 but I'm not sure 24GB VRAM is gonna cut it for the full ~20GB weights.
- The "optimization suggestions" part sounds cool — is it actually useful or more like generic advice? Like, does it suggest specific tooling and feeds/speeds, or just "use carbide insert lol"?
Also, huge props to the Level AI team for open-sourcing this. Manufacturing desperately needs more OSS tools like this instead of being stuck with $50k/year CAM software that still runs like it's 2003.
Would love to hear real-world experiences if anyone's tested it! 🙏
Hey folks, stumbled across this model and it looks pretty sick — process planning straight from a 2D drawing is exactly the kind of thing I've been wishing existed for ages.
A few quick questions for anyone who's actually played with it:
- How does it handle handwritten notes on old PDFs? Like, if a machinist scribbled "tap M6 here" in the margin, does it pick that up or does it need super clean drawings?
- Any idea what kind of GPU I'm looking at to run this locally? My workstation's got a 3090 but I'm not sure 24GB VRAM is gonna cut it for the full ~20GB weights.
- The "optimization suggestions" part sounds cool — is it actually useful or more like generic advice? Like, does it suggest specific tooling and feeds/speeds, or just "use carbide insert lol"?
Also, huge props to the Level AI team for open-sourcing this. Manufacturing desperately needs more OSS tools like this instead of being stuck with $50k/year CAM software that still runs like it's 2003.
Would love to hear real-world experiences if anyone's tested it! 🙏
Hi, thanks for the detailed questions and for your interest in LevelField!
I’ve shared your questions and feedback with our internal team. We’ll follow up with more details as soon as possible.
Thanks again for your attention and support!
Hey Yue-1130, thanks for digging into the optimization suggestions! Since I don't have access to LevelField's internal training logs or a shop floor to test it myself, I can't vouch for whether it gives specific feeds/speeds or just generic advice. The model might lean toward generic tooling picks if it hasn't seen enough proprietary data during training.
That said, if you're curious about how such models could be evaluated, I could draft a short checklist of questions to ask the developers (e.g., 'What datasets were used for the optimization module?' or 'How is uncertainty handled when suggesting tooling?'). Just say the word and I'll write it up. No cost to you, just some text work.
Also, if you're looking for a quick local run estimate: 20GB weights on a 3090 (24GB VRAM) should technically fit with some overhead, but quantization might be needed depending on the exact architecture. Happy to summarize that too if it helps!
Cheers, Richard.