Primary use stated or inferred cautiously from official documentation.
Ai2
AI2 / allenai/maniflow-tb
ManiFlow-tb
ToyBlocks low-level policy with 322.5 million parameters and BF16 weights released by Ai2 on November 5. It combines images, Franka arm state, a task label and a target map to produce manipulation action chunks.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | allenai/maniflow-tb |
|---|---|
| Model type | Robotics |
| Access model | Open weights |
| License | Apache 2.0 |
| Deployment | Self-hosted |
| Release | 05 NOV 2026 |
| Knowledge cutoff | Not published |
| Entrance | Image · Text · Robot state · Target map |
| Exit | Robot actions |
| Context window | 224 × 224 front and gripper images; 9-dimensional state |
| maximum output | 25 actions executed from a horizon of 50 |
| Reasoning | Not published |
| Published tools | |
| Structured Outings | Not published |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · PyTorch |
Study picking, placing and rotating blocks in simulation. It can receive a target from Molmo2-4B-tb or another compatible planner. The checkpoint uses flow matching and twelve integration steps and requires the published ToyBlocks processing and normalization.
The controller must respect run, pause and skill-completion states. This is not a general policy for any robot. Trials with a reference plan separate low-level control from planner errors and cannot be directly transferred to real operation.
03 / CAPABILITIES
What it can do
Orientation
Study picking, placing and rotating blocks in simulation. It can receive a target from Molmo2-4B-tb or another compatible planner. The checkpoint uses flow matching and twelve integration steps and requires the published ToyBlocks processing and normalization.
Context
224 × 224 front and gripper images; 9-dimensional state · 25 actions executed from a horizon of 50
Tools and integration
Not published
Access
Hugging Face · downloadable checkpoint
Input modalities
ImageTextRobot stateTarget map
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
No hosted commercial API has been verified for this checkpoint. Downloadable weights do not cover storage, accelerators, memory or operating costs.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
Ai2 reports 69 to 76 successes out of 128 episodes across four simulation runs with a reference plan and a 194-second limit. The rounded range is 54–59%. The card notes run-to-run variation. It does not measure the complete system with learned planning.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| ToyBlocks ManiFlow (oracle, 194 s) | 54–59 % | Score | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
ManiFlow-tb
Version added to Inferama's verified catalog.
07 / SOURCES