Harsha Gundala — GPC-1 general-purpose classifier with continuous outputs for bboxes, poses, coordinates, mechanical control

TL;DR

@harshagundal open-sources GPC-1 on Hugging Face at harshatheg/GPC-1, a Qwen-based “general-purpose classifier” that goes beyond label-picking to emit continuous numerical outputs — bounding boxes, poses, coordinates, exact angles/distances, and even direct mechanical control values — in millisecond latency. It is post-trained on “millions of datapoints and thousands of unique problem classes” mined from public X posts, and uses a dual training technique that jointly updates the Qwen base plus attached adapters, positioned as an easy-to-fine-tune starting point. Announced as a self-contained tweet with a 45-second video demo and follow-ups noting the fine-tuning-friendly design.

Key claims

GPC-1 introduces a new output class beyond categorical labels: “precise continuous range numerical outputs” covering bounding-box coordinates, exact angles and distances, pose estimations, and scores [tweet body].
Positioned as usable for direct mechanical control (“bounding boxes, poses, coordinates, direct mechanical control and so much more”) at millisecond latency [tweet body].
Post-trained on “millions of datapoints and thousands of unique problem classes learned directly from you guys posting on X” — public-X-derived supervision as the training signal [tweet body].
Uses a dual training technique that jointly trains the Qwen base model and add-on adapters, explicitly designed so downstream users can fine-tune it further with minimal effort [follow-up reply].
Released on Hugging Face at post time as harshatheg/GPC-1 [linked follow-up].

Method

The tweet is a launch announcement, not a technical writeup. What is disclosed: a Qwen base model plus adapters trained jointly (not adapter-only, not base-only) on a large corpus of public X posts covering thousands of problem classes, with an added output head or format that emits precise continuous ranges rather than one-of-N label logits. The reply thread emphasizes that the joint-adapter training is a deliberate choice to keep the model easy to specialize downstream. No architecture diagram, no parameter count, no training-corpus statistics beyond “millions of datapoints and thousands of unique problem classes,” no benchmark numbers, no comparison against Qwen-VL / LocateAnything / Perceptron Mk1 / other continuous-output VLMs.

Results

The single quantitative claim is “in milliseconds” per inference; no benchmark comparison, no calibration, no per-task accuracy. Evidence at launch is the attached 45-second video demo (bounding boxes / poses / coordinates presumably shown) and an “early user feedback is also good” line. Reply thread from @yoheinakajima notes his own local-VLM classifier project runs ~210ms p50 on Apple M5 with PyTorch/MPS FP16, providing a useful external latency reference point for the same “small VLM as always-on classifier” niche.

Why it’s interesting

This is the second harshagundal launch on the wiki after Qwen-2.5-1B-RLCD open-sourced as tongue-in-cheek response to System One / Jev launch (the tongue-in-cheek Qwen-2.5-1B-RLCD reply to the Introducing System One Models & Jev launch), and it lands in the same lane as several serious research efforts the wiki has already filed. The continuous-output-VLM-as-primary-interface thesis is exactly what LocateAnything: Fast and High-Quality Vision-Language Grounding with Parallel Box Decoding argues for at the box level (Parallel Box Decoding as a single atomic-unit output replacing coordinate-token serialization) and what Introducing Perceptron Mk1 pitches as a productized VLM sidecar (point / box / polygon / track / clip / 21-keypoint pose as first-class outputs). GPC-1’s contribution over those is the “and direct mechanical control” extension — using the same continuous-output head to close a robot loop instead of only labeling video — but the tweet ships zero measurements to back that claim, unlike LocateAnything (throughput + F1 numbers vs Qwen3-VL and Rex-Omni) or Perceptron Mk1 (priced product tier with cost benchmarks). The dual base+adapter training claim is worth flagging for the Parameter-Efficient Finetuning cluster: prior wiki entries on adapter fine-tuning (LoRA/PiSSA and successors) mostly freeze the base — GPC-1 explicitly does not — but again there’s no ablation showing this joint-training choice matters. Read charitably, this is a public existence proof that a small Qwen + adapters can serve as an always-on continuous-output classifier for robotics-adjacent primitives; read strictly, it’s an unbenchmarked HF release from an author whose previous entry the wiki flagged as ironic rebuttal rather than serious science.