Why Banking77 is hard
Noisy labels
Intent data is messy. Banking77 does not give you a clean synthetic path to good-looking results.
SeedFrontier.ai
Banking77 breakthrough
Intent classification result
Banking77 is harder than it looks.
Noisy labels. 77 intents. Real overlap. SeedFrontier just hit 94.42% on the official Banking77 test set, up 0.59 percentage points over baseline, under a strict full-train protocol with no leakage.
~225 ms. ~68 MiB. Small enough to run on-robot. Fast enough to keep a humanoid standing. This is the shape humanoid robotics actually needs.
94.42%
official Banking77 test accuracy
Measured on the official test set
+0.59pp
over baseline
Meaningful lift under the same evaluation frame
~225 ms
inference
Runtime kept visible instead of hand-waved away
~68 MiB
model footprint
Compact enough to matter in deployment discussions
Result snapshot
What matters
Accuracy
94.42%
official Banking77 test score
Delta
+0.59pp
improvement over baseline
Runtime
~225 ms
inference
Footprint
~68 MiB
model size
Difficulty profile77 intents
Label noisehigh
Intent overlaphigh
Production viabilitystrong
Why Banking77 is hard
77 intents
This is a large intent inventory with plenty of room for confusion across semantically adjacent classes.
Why Banking77 is hard
Real overlap
The hard part is not just class count. It is the genuine semantic overlap between user intents.
The headline
A benchmark result that reads like engineering, not marketing.
The point of this page is not to shout “state of the art” in the abstract. It is to show a result that is specific, measurable, and anchored to a known benchmark with visible operating characteristics.
94.42% on the official Banking77 test set, with a strict full-train protocol, no leakage, roughly 225 milliseconds inference, and a footprint around 68 MiB.
Protocol
Credibility comes from the rules around the number.
The landing page needs to make the evaluation discipline obvious. That is what separates a real benchmark statement from a dressed-up experiment.
Official test set
The headline number is the real external score, not a cherry-picked split or internal holdout.
Strict full-train protocol
The result was produced with a disciplined training setup rather than ad hoc experimentation.
No leakage
The page makes credibility explicit. No contamination, no hidden shortcut, no soft benchmark framing.
Measured runtime + size
Accuracy is paired with inference and memory footprint so the result reads like a deployable system.
Why this matters
The result is strong because the story is disciplined.
- This is a stronger public story than a vague AutoML claim because the benchmark, delta, and constraints are concrete.
- The result says SeedFrontier can push intent classification quality without disappearing into giant-model handwaving.
- The combination of score, protocol discipline, runtime, and footprint gives the page a credible engineering tone.
Messaging line
Banking77 is not easy. That is exactly why this result is worth publishing.
Robotics
This is the shape humanoid robotics actually needs.
Tesla Optimus. Boston Dynamics. Figure. 1X. Agility Robotics. Every humanoid and legged platform being built right now is compute-starved. A giant cloud-scale model does not fit in a robot torso powered by a battery. A ~68 MiB model running in ~225 ms does.
That is not a coincidence. It is the frontier these platforms actually live on — and it is where SeedFrontier is built to operate.
Compute is trapped on-robot
Humanoid platforms run on embedded accelerators with thermal and power budgets measured in watts. Cloud offload is not an option when control loops run at kilohertz rates.
Latency is standing vs. falling
Balance, locomotion, and manipulation policies run anywhere from 30 Hz to 1 kHz. Every millisecond of inference is a direct physical constraint on what the robot can actually do.
Battery budgets punish bloat
Every joule spent on inference is a joule not spent on actuators. Smaller, faster models extend runtime and make concurrent on-robot skills possible in the first place.
The on-robot control stack
Every layer is a deployment constraint.
A humanoid runs many models at once, at wildly different frequencies and sizes. Large general models cannot fill these layers. Small, specialized, production-shaped models can.
LayerFrequencySize budget
Balance and whole-body control500–1000 Hzunder 10 MB
Joint-level control1000+ Hzunder 1 MB
Locomotion policy50–200 Hz1–50 MB
Manipulation policy30–100 Hz50–500 MB
Visual perception30–60 Hz20–200 MB
Task planning (VLM)1–5 Hz2–7 B params
Who this is for
Humanoid and legged robotics teams shipping on real hardware.
If your platform has a thermal budget, a battery, and a control loop, the shape of your models is the shape of your product. Small, fast, accurate models are not a nice-to-have for humanoid robotics — they are the only thing that fits.
Tesla OptimusBoston DynamicsFigure1XAgility RoboticsApptronikSanctuary AIPhysical Intelligence
Messaging line
Humanoids do not need bigger models. They need the right model — small, fast, and accurate enough to run on-robot without apology. SeedFrontier delivers that shape.
Coming next
Full results soon on seedfrontier.ai
Banking77 is a proof point. The challenge is real, the benchmark is official, the gain is measurable, and the runtime profile stays visible — the kind of shape a real deployment actually needs.
The real target is humanoid robotics, where small, fast, specialized models are not optional. The deeper write-up will connect the Banking77 result to on-robot deployment and what this means for the teams building the next generation of humanoid platforms.