One model, every sense: omni-multimodal reaches robots too
Summary
Three labs shipped genuinely native multimodality in one late-July/early-August 2026 window -- not a vision head bolted onto a language model. Google DeepMind's Gemini Robotics 2 (July 30) drives three different robot embodiments, including a five-fingered 22-degree-of-freedom hand, from a single checkpoint, adapting to a new bi-arm robot in under 200 examples. Black Forest Labs' FLUX 3 (July 23) folds image, video and audio into one flow-matching model, with robot action finetuned on top and tested at Audi. And MiniMax open-weighted H3 (Aug 3), a 33B model that ingests text/image/video/audio and returns up-to-2K video with native 32 kHz stereo audio -- though its 'community' license excludes the US, EU, UK and South Korea.
The shared move is one weight set, many senses -- and, for two of the three, many outputs. Gemini Robotics 2 controls a gripper, an Apollo humanoid and a dexterous hand without swapping models; FLUX 3 and H3 both generate across modalities from a single backbone. What is not evenly distributed is reliability and openness -- the fine print, not the capability, is where these diverge.
View data as table
| Unscrew bulb (SharpaWave) | 92% | multi-finger dexterity, Apollo + SharpaWave hand |
|---|---|---|
| Precise insertion (Franka Duo) | 89.6% | gripper dexterity |
| Diverse tool kitting (Franka Duo) | 78.9% | gripper dexterity |
| Pick from shelf (Apollo Inspire) | 76.3% | whole-body manipulation |
| Pick from table (Apollo Inspire) | 68.4% | whole-body manipulation |
| Pick from floor (Apollo Inspire) | 45.7% | lowest reported rate |
The three, model by model
- Gemini Robotics 2 (Google DeepMind, July 30): one checkpoint drives three different robot embodiments, including the five-fingered, 22-degree-of-freedom SharpaWave hand on Apollo 2, and adapts to a new bi-arm robot in under 200 examples. Vendor-reported success ranges from 92% (unscrew a bulb) down to 45.7% (pick an object off the floor).
- FLUX 3 (Black Forest Labs, July 23): a single flow-matching model that jointly learns from images, videos and audio; robot action is a finetuned capability layered on its pretrained video backbone, now being tested on production dexterous-manipulation tasks at Audi. In preliminary video evals it was preferred over Runway Gen-4.5 in 77% of comparisons.
- MiniMax H3 (Aug 3): a 33B-parameter dense, single-stream omni Transformer that understands text, images, video and audio and generates 4-15s video at up to 2K/24fps with native 32 kHz stereo audio -- open weights on Hugging Face.
- The fine print: H3's weights are open but its MiniMax H3 Community License names the EU, UK, South Korea and the US as 'Excluded Territories', and requires written authorization for commercial use above $20M in yearly revenue -- restrictive, not permissive.
- The through-line: native multimodality -- one architecture, one weight set, many senses -- is now the baseline, including for robots. The differentiation has moved to embodiment breadth, output reliability, and license terms.
Corrections and caveats. FLUX 3's robot-action ability is a finetuned capability built on its video backbone, not a fourth jointly-trained modality, and Black Forest Labs frames the Audi work as dexterous manipulation on production tasks. All Gemini Robotics 2 and FLUX 3 figures here are vendor-reported and not independently replicated. Gemini Robotics 2's safety report is confirmed to resolve as an 18-page first-party PDF, but its compressed text was not machine-extracted, so no claim above rests on its interior. MiniMax H3's weights are open; its license is not.
Sources(5) ▾
- Google DeepMind, Gemini Robotics 2 brings whole-body intelligence to robots (2026-07-30) — Google DeepMind's own launch post for Gemini Robotics 2. Source for the July 30, 2026 release, the single-checkpoint control of three different robot embodiments, the five-fingered 22-degree-of-freedom SharpaWave hand on Apollo 2, the sub-200-example bi-arm adaptation claim, and the reported task success rates (unscrew bulb 92%, precise insertion 89.6%, diverse tool kitting 78.9%, pick from shelf 76.3%, table 68.4%, floor 45.7%). Fetched and read in full. deepmind.google
- Google DeepMind, Gemini Robotics 2: Safety report (2026-07-30) — The companion safety report linked from the Gemini Robotics 2 launch post. Confirmed to resolve as a linearized 18-page, ~10 MB first-party PDF hosted on DeepMind's media storage; the automated fetcher could not extract its compressed text, so no fact in this digest rests on its interior text -- it is cited as the issuer-hosted safety companion only. storage.googleapis.com
- Black Forest Labs, FLUX 3 (2026-07-23) — Black Forest Labs' own FLUX 3 announcement. Source for the July 23, 2026 Early Access availability, the unified flow-matching architecture that jointly learns from images, videos and audio, the finetuned-from-video-backbone action models, the Audi production-task testing, and the preliminary 77% preference over Runway Gen-4.5. Fetched and read in full. bfl.ai
- MiniMax AI (Hugging Face), MiniMax-H3 (model card) (2026-08-03) — Official MiniMax-H3 model card and open-weights release on Hugging Face. Source for the 33B-parameter dense single-stream H3-Omni-Transformer, the text/image/video/audio input understanding, the 4-15s / 24fps / up-to-2K video output with 32 kHz stereo audio, and the release under the MiniMax H3 Community License Agreement. Fetched and read in full. huggingface.co
- MiniMax AI (Hugging Face), MiniMax H3 Community License Agreement (2026-08-03) — The full text of the license the H3 weights ship under. Source for the 'Excluded Territories' (European Union, United Kingdom, Republic of Korea and United States) and the $20M-yearly-revenue threshold above which separate written authorization is required. Fetched and read in full. huggingface.co
Comments
Always open. Logged-in readers can annotate paragraphs in place.
Three labs shipped genuinely native multimodality in one window -- not a vision head bolted onto a language model. Google DeepMind's Gemini Robotics 2 (July 30) drives three different robot embodiments -- including a five-fingered, 22-degree-of-freedom hand -- from a single model checkpoint, and adapts to a new bi-arm robot in under 200 examples. Black Forest Labs' FLUX 3 (July 23) folds image, video and audio into one flow-matching model. And MiniMax open-weighted H3 (Aug 3), a 33B model that ingests text, image, video and audio and returns up-to-2K video with native 32 kHz stereo audio.