π₀
Physical Intelligence's Open Robot Foundation
Physical Intelligence's flagship vision-language-action model — π₀ (October 2024) opened the open-weight VLA category. π₀.₅ (April 2025) added open-world generalization; π*0.6 (Nov 2025) added experience-based RL. Released through the openpi GitHub repo, with reference deployments on commodity robot arms.
Speed
Medium
66 tok/s output
Cost
High
$30.00 in / $150.00 out
How are Intelligence, Speed & Cost bucketed?
Intelligence and Speed buckets are percentile ranks on
Artificial Analysis. Cost buckets are fixed dollar
thresholds keyed off output-token price ($/M out).
Intelligence
- Top 1%≤ 1%
- Top 5%≤ 5%
- Top 10%≤ 10%
- Good≤ 25%
- Medium≤ 50%
- Below avg> 50%
Speed
- Top 1%≥ 345 tok/s
- Top 5%≥ 237 tok/s
- Top 10%≥ 196 tok/s
- Good≥ 146 tok/s
- Medium≥ 90 tok/s
- Slow< 90 tok/s
Cost
- Freeopen weights · self-host
- Low< $1 / M out
- Moderate$1–5 / M out
- High≥ $5 / M out
Why it matters
Made open-weight VLA a real thing — before π₀, every serious VLA was closed (Google's RT, Tesla's, Figure's). Pi proved you could open-source the architecture and still build a business on hardware + integration + commercial fine-tuning.
Core Capabilities
Agent Workflows
Built for tool use and autonomous tasks.
Multimodal
Combines text, vision, and audio in one model.
Generative
Produces images, video, audio, or other media.
Vision
Understands images, scenes, and visual context.
Context Window
Context window not disclosed.
Availability
API
Not available
Product / App
Not available
Open Source
Released
Enterprise
—
Pricing Model
Free / self-host
Open weights — pay only for compute.
Self-host Capability / Performance
Where this model sits relative to the middle 60% of models in the tree. All scores are 0–10 (higher is better).
Lower 20% Upper 80% This model
Lower 20% 20th percentile — 20% of models score below this This model Where the current model lands Upper 80% 80th percentile — only 20% of models score above this
Percentile boundaries are computed across every model in the tree that reports the underlying benchmark for each capability.
What it feels like
- First open-weight Vision-Language-Action (VLA) model that genuinely competes with closed Google RT / Tesla / Figure stacks
- Flow-matching action head (instead of autoregressive tokens) — generates continuous robot trajectories with physical smoothness
- Folds laundry, busses dishes, packs groceries — the demo videos that convinced the field VLAs were real
- 3B parameters — small enough to run on edge robot compute, not in a datacenter
- openpi GitHub repo lets any lab fine-tune for their gripper / arm / workspace
- Successors (π₀.₅ open-world; π*0.6 experience-based RL) keep extending the open VLA frontier
Best use cases
- Robotics research and academic VLA work
- Startups building physical-AI products who need a foundation to fine-tune from
- Long-horizon manipulation tasks (multi-step household / warehouse routines)
- Distillation into smaller task-specific robot policies
Not ideal for
- Pure language reasoning, coding, or chat — π₀ is a robot policy, not an LLM
- Tasks requiring out-of-distribution objects or scenes far from training data (improved in π₀.₅)
- Real-time loops below ~10 Hz on commodity robot CPUs (π₀-FAST helps but still constrained)