9B-Parameter Model Beats GPT-6 on Six of Eight Spatial Benchmarks — But the Two Losses Go Unmentioned
This morning, Daxiao Robotics (大晓机器人), a SenseTime-affiliated startup, released its embodied foundation model ACE-Brain-0.6 with a loud promotional push: on eight spatial anchoring benchmarks, it claims victory on six against OpenAI's GPT-6 Astra.
The scorecard: Where2Place 76.55 vs 71.10, PixMo-Count 89.41 vs 88.10, RoboSpatial 74.82 vs 69.55, RoboAfford 88.51 vs 84.94, ShareRobot 49.70 vs 46.60, RobotGrounding 94.13 vs 91.10. A model one nine-hundredth the size, leading the industry flagship on six benchmarks.
Bottom line up front: these numbers have real substance, but there are two things the announcement didn't proactively tell you. Let's break it down one by one.
What Kind of Contest Is This?
ACE-Brain-0.6 is a specialist. It uses Qwen3.5-9B as its base, freezes the vision encoder, and fine-tunes on roughly one million supervision samples aimed squarely at "spatial grounding." Spatial grounding sounds abstract, but it's concrete: to understand "put the fruit in the basket," a robot must first know where the basket's edge is, where the free space is, and which part of the object to grasp. General models must chat, write code, play chess, and diagnose illness; this model trains on exactly one thing — looking at images and finding positions.
A specialist beating a decathlete on one event is no surprise. The press release itself admits this in the body text — "specialized capability breakthrough" — even if the headline buried the word "specialized."
The first thing not said: eight benchmarks, six wins — which two did it lose? Not a single word. The average across seven public benchmarks is 74.53 vs 73.68, a gap of just 0.85 points. The prettiest wins were Where2Place and RoboSpatial (roughly five-point leads), while on ShareRobot both models are stuck in the forties — the hardest benchmark, which nobody passes. Listing the wins item by item while leaving the losses unnamed is a release posture worth remembering.
The second thing: the eighth benchmark, RobotGrounding, was built by Daxiao itself — 2,591 questions across 14 task types covering object localization, small-object recognition, and cross-camera correspondence. Writing your own exam invites home-field suspicion. But one detail cuts the other way: GPT-6 Astra scored 91.10 on this unfamiliar test. If the grading were rigged, a general model couldn't reach ninety-one. The test didn't trip Astra up — it just didn't help it.
How the Million Samples Were Made
The real value of this release isn't the scorecard — it's the data pipeline, because it's reusable.
Robots record enormous volumes of video daily: gripper states, action trajectories, task instructions. But these record *what was done*, not *why it grabbed here* or *where it should place things*. Daxiao's pipeline takes 30-plus open-source robot data families plus private egocentric data through an automated process: filter keyframes by gripper state and motion magnitude, automatically build scene object vocabularies, use Grounding DINO and SAM 3 for object localization and region identification, then generate multiple QA types — target referring, counting, spatial relations, and "target absent."
Two design choices stand out. First, "target absent" questions were deliberately constructed — ask the robot for something not in the scene, and the correct answer is "not there," not a plausible-sounding fabrication. A robot must know not only what to act on, but when *not* to act. Second, every answer carries verification: counting questions are generated only when object instance annotations are complete, and "target absent" questions must confirm the object truly isn't in frame. Roughly 940,000 images and 2.7 million QA pairs were compressed into about one million training samples.
This "verifiable" approach is the hardest part of the data-quality promise.
The Real Finding: Adding Coverage Beats Adding Data
The most weighty experiment is a controlled comparison: architecture and training method held completely fixed, with data scaled from 260K to one million samples.
The seven-public-benchmark average rose from 69.3 to 74.5. Growth is approximately log-linear: each doubling of data yields about 2.4 points. Going from 260K to one million is nearly four doublings, predicting 4.7 points — actual gain was 5.2, with the curve even bending slightly upward.
But the other half is the point: simply piling up more of the *same* data gains less than one point. What drives growth is continuously adding new scenes, new spatial relations, new supervision forms. In other words, the marginal value of a second batch of similar data is near zero; the first batch of new-scene data is priceless. For every "we have 100 million samples" story, this is a reminder: data value lies in boundary expansion of coverage, not depth of accumulation.
Where This Company Stands
Daxiao Robotics was founded in mid-2025. Chairman Wang Xiaogang is a SenseTime co-founder who earlier led the DeepID series and SenseTime's autonomous driving efforts; the company raised several hundred million USD in cumulative funding in the first half of this year. The ACE-Brain series shipped three generations in five months: v0 in March (with Shanghai Jiao Tong, NTU Singapore, CUHK, and HKU, refreshing 19 SOTA results), open-sourced v0.5 in July (a unified cognitive loop of perception, planning, action, evaluation, and evolution), and today's 0.6 pivoting to specialist breakthrough.
The open-source tradition is real: v0 and v0.5 code and weights are on GitHub and HuggingFace. The 0.6 weights had not been uploaded at press time — the release is only hours old. Once they land, third parties can recompute all six comparison numbers. Until then, today's scorecard is: officially self-reported, reprinted from one source, pending verification.
One-line takeaway: a 9-billion-parameter model with one million specialized samples beating a much larger general model on six tasks — the biggest lesson isn't "small models strike back," but that coverage breadth of specialized data is worth more than depth of accumulation. The 2.4-points-per-doubling curve and the sub-1-point same-type-data control are the two numbers most worth taking away from this release.
---
*Sources: Daxiao Robotics press release (PingWest reprint 2026-10-19; Sohu same day); ACE-Brain-0 technical report arXiv 2603.03198 (2026-03); ACE-Brain-0.5 arXiv 2607.04426 (2026-07); GitHub DAXIAORobotics/ACE-Brain-0 (86★), ACE-Brain-0.5 (39★); HuggingFace ACERobotics 8B weights; Caixin 2026-06-15 (SenseTime affiliation, funding); Science and Technology Daily 2026-07-07 (0.5 open source). The 0.6 technical report and weights were not available at press time.*