Ai2Specialist modelOpen weights
ManiFlow-tb
Summary
Ai2 released ManiFlow-tb on November 6, 2026 as a low-level action model for ToyBlocks. It controls a simulated or physical Franka robot arm for tabletop block tasks, including picking, placing, and rotating. A high-level planner supplies the skill and target.
The model processes front and gripper RGB images, robot state, task name, and a target feature from the higher-level model. Its ManiFlowMultiModal architecture with a DiTX action head has 322.5 million parameters. It outputs chunks of 25 executed actions with a prediction horizon of 50.
The BF16 Safetensors weights and training code use Apache 2.0. Ai2 evaluated the model in ToyBlocks on 128 held-out episodes, using an oracle as the high-level planner and a 194-second time limit. Four runs reached 54% to 59% task success. The simulator and batched inference are not fully deterministic.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 4 published benchmark results
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
ToyBlocks ManiFlow (run 1) (Show measurement, test conditions, and source)
- Source value
- 57
- Score
- 57
- Metric
- task success rate
- Unit
- %
- Category
- robotics
- Direction
- Higher is better
- Sample
- 128
- Source type
- vendor-reported
- Evaluator
- Ai2 ManiFlow-tb model card
- Status
- active
- Retrieved at
- 2026-11-07
- Methodology
- Vendor-reported result. Simulation with a recorded high-level plan and a 194 s time limit. Run 1 on seven GPUs, 73 of 128 tasks successful. All runs use the same tasks.
ToyBlocks ManiFlow (run 2) (Show measurement, test conditions, and source)
- Source value
- 59
- Score
- 59
- Metric
- task success rate
- Unit
- %
- Category
- robotics
- Direction
- Higher is better
- Sample
- 128
- Source type
- vendor-reported
- Evaluator
- Ai2 ManiFlow-tb model card
- Status
- active
- Retrieved at
- 2026-11-07
- Methodology
- Vendor-reported result. Simulation with a recorded high-level plan and a 194 s time limit. Run 2 on seven GPUs, 75 of 128 tasks successful. All runs use the same tasks.
ToyBlocks ManiFlow (run 3) (Show measurement, test conditions, and source)
- Source value
- 59
- Score
- 59
- Metric
- task success rate
- Unit
- %
- Category
- robotics
- Direction
- Higher is better
- Sample
- 128
- Source type
- vendor-reported
- Evaluator
- Ai2 ManiFlow-tb model card
- Status
- active
- Retrieved at
- 2026-11-07
- Methodology
- Vendor-reported result. Simulation with a recorded high-level plan and a 194 s time limit. Run 3 on one GPU, 76 of 128 tasks successful. All runs use the same tasks.
ToyBlocks ManiFlow (run 4) (Show measurement, test conditions, and source)
- Source value
- 54
- Score
- 54
- Metric
- task success rate
- Unit
- %
- Category
- robotics
- Direction
- Higher is better
- Sample
- 128
- Source type
- vendor-reported
- Evaluator
- Ai2 ManiFlow-tb model card
- Status
- active
- Retrieved at
- 2026-11-07
- Methodology
- Vendor-reported result. Simulation with a recorded high-level plan and a 194 s time limit. Run 4 on seven GPUs, 69 of 128 tasks successful. All runs use the same tasks.
Similar models
More models of the same kind as ManiFlow-tb, chosen by model line, developer, size, price and release date.
- Molmo2-4B-tbAi2 · same developer
- LFM2.5-VL-3B-DSparkLiquid AI · similar size
- EmbeddingGemma 2Google DeepMind · similar size
- pplx-embed-v2-context-0.6bPerplexity · similar size
- pplx-embed-v2-late-0.6bPerplexity · similar size
- Falcon-OCR-ArabicTechnology Innovation Institute · similar size
- TeleOCRTeleAI · released around the same time
- Cohere Rerank 5 FastCohere · similar size
All specifications
Architecture
- Architecture
- ManiFlowMultiModal with a DiTX action head, 12 layers, 8 heads, and 322.5 million parameters. Weights are BF16 Safetensors.
- Source: Architecture
Watermarking and provenance
AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.
- AI watermarking
- Unknown
- no source
- Watermarking technique
- Not known
- no source
Licence and access
- Shutdown
- None announced
- no source
Further details
- Role
- Low-level learned action policy that executes movements for individual ToyBlocks skills. It complements a high-level planner such as Molmo2-4B-tb.
- Source: Role
- Input and output
- Inputs include front and gripper camera images, 9-dimensional robot states, task names, and target features. The output chunks 25 executed actions with a prediction horizon of 50.
- Source: Input and output
Sources and data date
Benchmark sources at a glance
Which independent evaluators tested this model. This is coverage, not a quality ranking.
No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.
Without vendor claims and historical or disputed results; those stay labelled in the single values.
| Type | Evidence and data date |
|---|---|
| Research status | Research date November 7, 2026. 1 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Additional source | Ai2 ManiFlow-tb model card (retrieved November 7, 2026) · Ai2 ManiFlow-tb model card · Editorial description reviewed November 7, 2026 |
| Additional source | Ai2 ToyBlocks release (retrieved November 7, 2026) · Ai2 ToyBlocks release · Editorial description reviewed November 7, 2026 |
| Additional source | Ai2 ManiFlow-tb model card (retrieved November 7, 2026) |
| Additional source | Ai2 Molmo2-4B-tb model card · Editorial description reviewed November 7, 2026 |