Black Forest LabsSpecialist modelOpen weights
FLUX 3 Action
PopularityRank 639 of 645
6out of 100first value
based on two data groups, mostly search
Summary
FLUX 3 Action is a 7-billion-parameter world action model for robotics. It takes camera frames, a state vector, and a text instruction. It then returns 32 actions and optionally 32 predicted video frames. Image and action tokens are denoised together. The open weights can be adapted to different robot systems.
The released base checkpoint is not a finished robot policy. A specific task still needs fine-tuning, for example on teleoperation data for a robot arm. The RoboLab-120 rate belongs to a DROID-fine-tuned policy, not the base checkpoint.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 26 published benchmark results
Page 1 of 3
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.
RoboLab-120 (Show measurement, test conditions, and source)
- Source value
- 17.29
- Score
- 17.29
- Metric
- RoboLab success rate
- Unit
- %
- Category
- robotics-action
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Black Forest Labs
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. Figure 4 native EE50 action head on the FLUX 3 Action 724K action-midtraining checkpoint. BFL discusses DROID fine-tuning separately in Figure 8. The BFL model card and pinned BFL setup/config identify the shared flux-3-action-base repository and distinguish the DROID-finetuned policy and its fresh DROID heads. No immutable digest for the evaluated 724K checkpoint is published.
- Context
- BFL does not publish per-result episode count or a confidence interval. The result is a checkpoint-scoped preliminary native-EE50 evaluation, not an unqualified profile-wide score. The evaluated 724K artifact is not bound to a captured immutable weights digest. The source does not identify a benchmark-suite version.
FLUX 3 Action RoboLab success-rate mean: Base BF16 (Show measurement, test conditions, and source)
- Source value
- 41.6
- Score
- 41.6
- Metric
- RoboLab success rate (%)
- Unit
- %
- Category
- robotics-action
- Direction
- Higher is better
- Sample
- 6
- Source type
- vendor-reported
- Evaluator
- Black Forest Labs FLUX 3 Action research page, Table 5
- Status
- disputed
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. Exact BFL Table 5 row "Base BF16" for FLUX 3 Action (F3A). Table caption identifies RoboLab success rates for F3 base, guidance-distilled, and step-distilled policies under BF16 and FP8 serving. The table's Seeds column lists N=6; the reported mean is across the listed seeds. No immutable checkpoint digest or benchmark-suite version is attached to the row.
- Context
- BFL reports mean ± SEM, not a confidence interval; no CI fields are populated. The source table's N is the number of seeds, not task episodes. A separate unqualified Table 1 headline gives FLUX 3 Action 42.92%, numerically equal to one listed Base BF16 seed; it does not document the same aggregation/configuration as this Table 5 mean. The Base BF16 row is therefore marked disputed rather than treated as an uncontested active aggregate. The frozen baseline's row research:flux-3-action:0 is 17.29% for a different native EE50 724K midtraining checkpoint and is not replaced or reconciled here.
FLUX 3 Action RoboLab success-rate mean: Base FP8 (Show measurement, test conditions, and source)
- Source value
- 41.16
- Score
- 41.16
- Metric
- RoboLab success rate (%)
- Unit
- %
- Category
- robotics-action
- Direction
- Higher is better
- Sample
- 8
- Source type
- vendor-reported
- Evaluator
- Black Forest Labs FLUX 3 Action research page, Table 5
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. Exact BFL Table 5 row "Base FP8" for FLUX 3 Action (F3A). Table caption identifies RoboLab success rates for F3 base, guidance-distilled, and step-distilled policies under BF16 and FP8 serving. The table's Seeds column lists N=8; the reported mean is across the listed seeds. No immutable checkpoint digest or benchmark-suite version is attached to the row.
- Context
- BFL reports mean ± SEM, not a confidence interval; no CI fields are populated. The source table's N is the number of seeds, not task episodes. A separate unqualified Table 1 headline gives FLUX 3 Action 42.92%, numerically equal to one listed Base BF16 seed; it does not document the same aggregation/configuration as this Table 5 mean. The Base BF16 row is therefore marked disputed rather than treated as an uncontested active aggregate. The frozen baseline's row research:flux-3-action:0 is 17.29% for a different native EE50 724K midtraining checkpoint and is not replaced or reconciled here.
FLUX 3 Action RoboLab success-rate mean: Guidance Distilled BF16 (Show measurement, test conditions, and source)
- Source value
- 42.19
- Score
- 42.19
- Metric
- RoboLab success rate (%)
- Unit
- %
- Category
- robotics-action
- Direction
- Higher is better
- Sample
- 6
- Source type
- vendor-reported
- Evaluator
- Black Forest Labs FLUX 3 Action research page, Table 5
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. Exact BFL Table 5 row "Guidance Distilled BF16" for FLUX 3 Action (F3A). Table caption identifies RoboLab success rates for F3 base, guidance-distilled, and step-distilled policies under BF16 and FP8 serving. The table's Seeds column lists N=6; the reported mean is across the listed seeds. No immutable checkpoint digest or benchmark-suite version is attached to the row.
- Context
- BFL reports mean ± SEM, not a confidence interval; no CI fields are populated. The source table's N is the number of seeds, not task episodes. A separate unqualified Table 1 headline gives FLUX 3 Action 42.92%, numerically equal to one listed Base BF16 seed; it does not document the same aggregation/configuration as this Table 5 mean. The Base BF16 row is therefore marked disputed rather than treated as an uncontested active aggregate. The frozen baseline's row research:flux-3-action:0 is 17.29% for a different native EE50 724K midtraining checkpoint and is not replaced or reconciled here.
FLUX 3 Action RoboLab success-rate mean: Guidance Distilled FP8 (Show measurement, test conditions, and source)
- Source value
- 42.24
- Score
- 42.24
- Metric
- RoboLab success rate (%)
- Unit
- %
- Category
- robotics-action
- Direction
- Higher is better
- Sample
- 6
- Source type
- vendor-reported
- Evaluator
- Black Forest Labs FLUX 3 Action research page, Table 5
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. Exact BFL Table 5 row "Guidance Distilled FP8" for FLUX 3 Action (F3A). Table caption identifies RoboLab success rates for F3 base, guidance-distilled, and step-distilled policies under BF16 and FP8 serving. The table's Seeds column lists N=6; the reported mean is across the listed seeds. No immutable checkpoint digest or benchmark-suite version is attached to the row.
- Context
- BFL reports mean ± SEM, not a confidence interval; no CI fields are populated. The source table's N is the number of seeds, not task episodes. A separate unqualified Table 1 headline gives FLUX 3 Action 42.92%, numerically equal to one listed Base BF16 seed; it does not document the same aggregation/configuration as this Table 5 mean. The Base BF16 row is therefore marked disputed rather than treated as an uncontested active aggregate. The frozen baseline's row research:flux-3-action:0 is 17.29% for a different native EE50 724K midtraining checkpoint and is not replaced or reconciled here.
FLUX 3 Action RoboLab success-rate mean: Step Distilled BF16 (Show measurement, test conditions, and source)
- Source value
- 38.68
- Score
- 38.68
- Metric
- RoboLab success rate (%)
- Unit
- %
- Category
- robotics-action
- Direction
- Higher is better
- Sample
- 6
- Source type
- vendor-reported
- Evaluator
- Black Forest Labs FLUX 3 Action research page, Table 5
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. Exact BFL Table 5 row "Step Distilled BF16" for FLUX 3 Action (F3A). Table caption identifies RoboLab success rates for F3 base, guidance-distilled, and step-distilled policies under BF16 and FP8 serving. The table's Seeds column lists N=6; the reported mean is across the listed seeds. No immutable checkpoint digest or benchmark-suite version is attached to the row.
- Context
- BFL reports mean ± SEM, not a confidence interval; no CI fields are populated. The source table's N is the number of seeds, not task episodes. A separate unqualified Table 1 headline gives FLUX 3 Action 42.92%, numerically equal to one listed Base BF16 seed; it does not document the same aggregation/configuration as this Table 5 mean. The Base BF16 row is therefore marked disputed rather than treated as an uncontested active aggregate. The frozen baseline's row research:flux-3-action:0 is 17.29% for a different native EE50 724K midtraining checkpoint and is not replaced or reconciled here.
FLUX 3 Action RoboLab success-rate mean: Step Distilled FP8 (Show measurement, test conditions, and source)
- Source value
- 37.92
- Score
- 37.92
- Metric
- RoboLab success rate (%)
- Unit
- %
- Category
- robotics-action
- Direction
- Higher is better
- Sample
- 6
- Source type
- vendor-reported
- Evaluator
- Black Forest Labs FLUX 3 Action research page, Table 5
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. Exact BFL Table 5 row "Step Distilled FP8" for FLUX 3 Action (F3A). Table caption identifies RoboLab success rates for F3 base, guidance-distilled, and step-distilled policies under BF16 and FP8 serving. The table's Seeds column lists N=6; the reported mean is across the listed seeds. No immutable checkpoint digest or benchmark-suite version is attached to the row.
- Context
- BFL reports mean ± SEM, not a confidence interval; no CI fields are populated. The source table's N is the number of seeds, not task episodes. A separate unqualified Table 1 headline gives FLUX 3 Action 42.92%, numerically equal to one listed Base BF16 seed; it does not document the same aggregation/configuration as this Table 5 mean. The Base BF16 row is therefore marked disputed rather than treated as an uncontested active aggregate. The frozen baseline's row research:flux-3-action:0 is 17.29% for a different native EE50 724K midtraining checkpoint and is not replaced or reconciled here.
NVIDIA RoboLab-120 (Show measurement, test conditions, and source)
- Source value
- 42.9
- Score
- 42.9
- Metric
- success rate, overall (default language specificity)
- Unit
- %
- Benchmark version
- RoboLab-120
- Category
- robot manipulation
- Direction
- Higher is better
- Sample
- 1200
- Source type
- official-leaderboard
- Evaluator
- NVIDIA RoboLab-120 leaderboard
- Status
- disputed
- Retrieved at
- 2026-10-17
- Methodology
- Official benchmark leaderboard. Leaderboard model label FLUX 3 Action (WAM, 7B), RoboLab-120 Overall table, Default language-specificity variant. The source reports 515/1200 and SR%=42.9%; its episode-total data gives 1200 actual episodes for this policy and variant. The exact immutable evaluated checkpoint digest and full inference settings are not published.
- Context
- This is an additional NVIDIA-published leaderboard cell, not an established independent replication. The frozen preceding research note records a BFL Table 1 headline of 42.92% and says it matches one of the six Base BF16 seeds in BFL Table 5; the source does not crosswalk that value to a specific checkpoint, precision, or run. 515/1200 rounds to 42.9%, so the NVIDIA cell could describe the same underlying run or result. It is retained as disputed and must not be treated as an active separate score until that identity is resolved. It belongs to the same 1,200-episode evaluation as the difficulty and competency-axis breakdowns; those subgroups are not independent of the aggregate. The table header identifies a 95% Clopper-Pearson CI, but the captured row does not publish numeric bounds. The exact immutable checkpoint digest and complete inference configuration are not stated.
NVIDIA RoboLab-120 (Show measurement, test conditions, and source)
- Source value
- 49.1
- Score
- 49.1
- Metric
- success rate, simple
- Unit
- %
- Benchmark version
- RoboLab-120
- Category
- robot manipulation
- Direction
- Higher is better
- Sample
- 640
- Source type
- official-leaderboard
- Evaluator
- NVIDIA RoboLab-120 leaderboard
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Official benchmark leaderboard. Leaderboard model label FLUX 3 Action (WAM, 7B), RoboLab-120, Simple subgroup. The source cell reports difficulty simple, 314/640, 49.1%; total attempts are the denominator.
- Context
- The source reports successful attempts divided by attempts, not necessarily unique tasks. The difficulty groups sum to 1,200 attempts; the three competency-axis denominators sum to 1,480, so those axis groups overlap and must not be added. These are subgroups from one RoboLab-120 leaderboard run, not six independent evaluations. Exact checkpoint, precision and inference settings are not stated.
NVIDIA RoboLab-120 (Show measurement, test conditions, and source)
- Source value
- 39.2
- Score
- 39.2
- Metric
- success rate, moderate
- Unit
- %
- Benchmark version
- RoboLab-120
- Category
- robot manipulation
- Direction
- Higher is better
- Sample
- 390
- Source type
- official-leaderboard
- Evaluator
- NVIDIA RoboLab-120 leaderboard
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Official benchmark leaderboard. Leaderboard model label FLUX 3 Action (WAM, 7B), RoboLab-120, Moderate subgroup. The source cell reports difficulty moderate, 153/390, 39.2%; total attempts are the denominator.
- Context
- The source reports successful attempts divided by attempts, not necessarily unique tasks. The difficulty groups sum to 1,200 attempts; the three competency-axis denominators sum to 1,480, so those axis groups overlap and must not be added. These are subgroups from one RoboLab-120 leaderboard run, not six independent evaluations. Exact checkpoint, precision and inference settings are not stated.
NVIDIA RoboLab-120 (Show measurement, test conditions, and source)
- Source value
- 28.2
- Score
- 28.2
- Metric
- success rate, complex
- Unit
- %
- Benchmark version
- RoboLab-120
- Category
- robot manipulation
- Direction
- Higher is better
- Sample
- 170
- Source type
- official-leaderboard
- Evaluator
- NVIDIA RoboLab-120 leaderboard
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Official benchmark leaderboard. Leaderboard model label FLUX 3 Action (WAM, 7B), RoboLab-120, Complex subgroup. The source cell reports difficulty complex, 48/170, 28.2%; total attempts are the denominator.
- Context
- The source reports successful attempts divided by attempts, not necessarily unique tasks. The difficulty groups sum to 1,200 attempts; the three competency-axis denominators sum to 1,480, so those axis groups overlap and must not be added. These are subgroups from one RoboLab-120 leaderboard run, not six independent evaluations. Exact checkpoint, precision and inference settings are not stated.
NVIDIA RoboLab-120 (Show measurement, test conditions, and source)
- Source value
- 49.5
- Score
- 49.5
- Metric
- success rate, relational competency axis
- Unit
- %
- Benchmark version
- RoboLab-120
- Category
- robot manipulation
- Direction
- Higher is better
- Sample
- 420
- Source type
- official-leaderboard
- Evaluator
- NVIDIA RoboLab-120 leaderboard
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Official benchmark leaderboard. Leaderboard model label FLUX 3 Action (WAM, 7B), RoboLab-120, Relational competency axis subgroup. The source cell reports relational axis, 208/420, 49.5%; total attempts are the denominator.
- Context
- The source reports successful attempts divided by attempts, not necessarily unique tasks. The difficulty groups sum to 1,200 attempts; the three competency-axis denominators sum to 1,480, so those axis groups overlap and must not be added. These are subgroups from one RoboLab-120 leaderboard run, not six independent evaluations. Exact checkpoint, precision and inference settings are not stated.
Page 1 of 3
Popularity compared
How often the model is searched for, used and mentioned, compared with every model of its category.
Popularity index
0 to 100, higher is more popular
15 of 15 models
As of October 2026 · Source: gradually.ai popularity index
Over time
Weekly values from October 18, 2026 to November 9, 2026
FLUX 3 Actionpplx-embed-v2-context-9b-preview 11Praxis-1 1Gemini 3.8 Flash Cyber 0GPT-5.6 Cyber 0
Method changes in the timeline, a jump at these points is no market event:
- 1November 1, 2026: Value only with search or usage data, 30-day attention
- 2November 1, 2026: Method version 4 (common scale, shrinkage)
- 3November 9, 2026: New topics measurement V20 (daily quorum, short forms)
What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.
Similar models
More models of the same kind as FLUX 3 Action, chosen by model line, developer, size, price and release date.
- AstaBrief-8BAllen Institute for AI · similar size
- pplx-embed-v2-late-9bPerplexity · similar size
- pplx-embed-v2-context-9bPerplexity · similar size
- Qwen-Drive-1.0-4BQwen · similar size
- Molmo2-4B-tbAi2 · similar size
- Unlimited OCR 2Baidu · similar size
- LensVLM-9BApple · similar size
- TeleOCRTeleAI · released around the same time
All specifications
Architecture
Input and output
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
Sources and data date
Show all 14 sources
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 October 15, 2026. 7 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Unresolved | bare-profile-quality-benchmark Note Reported RoboLab-120 score 42.92 belongs to the separate DROID fine-tuned checkpoint, so it is not assigned to the generic profile. |
| Unresolved | flux-3-action-generic-checkpoint-score Note The source separately reports a RoboLab-120 score of 17.29% for the native EE50 checkpoint after 724K action-midtraining steps. The 42.92 score remains tied to the distinct DROID-100/DROID-Action fine-tuned checkpoint and is not assigned to the generic profile. The native EE50 row has no published scored-episode count, confidence interval or immutable checkpoint digest, so it is not a general result for an unspecified FLUX 3 Action checkpoint. |
| Not published | robolab-sample-count-and-confidence-interval Original wording The BFL Figure 4 and its surrounding text do not publish a scored episode count or confidence interval for the 17.29% native-EE50 checkpoint result. |
| Not published | evaluated-724k-checkpoint-digest Original wording The source identifies the checkpoint by its 724K action-midtraining step and native EE50 head but does not publish an immutable digest for the evaluated artifact. |
| Not published | checkpoint-digest-and-suite-version Note BFL publishes row-level variant and serving precision, but no immutable weights digest or RoboLab suite version in Table 5. |
| Unresolved | table-1-to-table-5-crosswalk Note The 42.92% Table 1 headline is unqualified and equals one Base BF16 seed, not the Table 5 Base BF16 mean; no source crosswalk establishes whether it uses the same settings/aggregation. |
| Not published | Exact single-step headline checkpoint and evaluation sample count Note BFL's current page text does not identify the exact distilled headline checkpoint, hardware/precision, or number of evaluation runs. |
| Additional source | Black Forest Labs announcement (retrieved September 26, 2026; October 16, 2026) · Black Forest Labs announcement · Editorial description reviewed October 2, 2026 |
| Additional source | FLUX 3 Action weights (retrieved September 26, 2026) · FLUX 3 Action weights · Editorial description reviewed October 2, 2026 |
| Additional source | Black Forest Labs (retrieved October 17, 2026; October 19, 2026; October 28, 2026) |
| Additional source | NVIDIA RoboLab-120 leaderboard (retrieved October 17, 2026) |
| Additional source | Model metadata (retrieved October 17, 2026)1234 |
| Additional source | OpenRouter: Popularity index (derived), CC BY 4.0 |
| Additional source | Vercel AI Gateway: Popularity index (derived), CC BY 4.0 |
| Additional source | Ollama: cumulative downloads for the popularity index |
| Additional source | Replicate: cumulative runs for the popularity index |
| Additional source | Google Trending Now: limited feed selection for Germany and the US |
| Additional source | Creative Commons Attribution 4.0 |