Runway AI Release Notes
180 release notes curated from 35 sources by the Releasebot Team. Last updated: Nov 19, 2026
Runway AI Products
- Nov 18, 2026
- Date parsed from source:Nov 18, 2026
- First seen by Releasebot:Nov 19, 2026
Request tags
Runway API adds request tags and tag-based credit usage breakdowns for generation requests, workflow runs, and the Developer Portal.
November 18th, 2026 - Generation requests can now carry tags in the X-Runway-Tags header, and POST /v1/organization/usage can filter and break credit usage down by those tags. A tag is a label such as campaign-q4 or a key:value pair such as team:growth. A Workflow run copies its tags onto every generation it starts. The same breakdown is on the Usage page in the Developer Portal. See Request tags and the Organization API reference to get started.
Original source - Nov 12, 2026
- Date parsed from source:Nov 12, 2026
- First seen by Releasebot:Nov 17, 2026
- Modified by Releasebot:Nov 18, 2026
Seedream 5.5 and Seedream 5.5 Layers on Runway Dev
Runway API adds Seedream 5.5 image generation and editing, with support for prompts and reference images, plus Seedream 5.5 Layers to split source images into editable transparent layers. Both models are available through the text-to-image endpoint, with flexible resolution and output options.
Seedream 5.5 and Seedream 5.5 Layers
November 12th, 2026 - Seedream 5.5 (seedream5.5) and Seedream 5.5 Layers (seedream5.5_layers) are now available on Runway Dev, both on the text to image endpoint.
Seedream 5.5 generates and edits images from a prompt of up to 4,000 characters, with up to 10 reference images and outputCount from 1–4. ratio is a width:height size across 1K, 1.5K, and 2K covering eight aspect ratios (1:1, 4:3, 3:4, 16:9, 9:16, 3:2, 2:3, 21:9), or auto_1k, auto_1.5k, or auto_2k to let the model choose the aspect ratio. outputFormat is png (the default) or jpeg, and background can be transparent (PNG only) or opaque. Billed at 2 credits per image at 1K or 1.5K and 4 credits at 2K, plus 0.8 credits per reference image. See image input requirements.
Seedream 5.5 Layers splits one PNG or JPEG source image into a base image and editable transparent PNG layers, with an optional prompt, a resolution of auto, 1k, 1.5k, or 2k, and maxImages from 2–100 (default 30) to cap the images returned. Billed per returned image at the same rates, with auto at the 2K rate: a decomposition reserves credits for maxImages, then settles to the images returned. See layer outputs and pricing.
Original source All of your release notes in one feed
Join Releasebot and get updates from Runway AI and hundreds of other software products.
- Nov 11, 2026
- Date parsed from source:Nov 11, 2026
- First seen by Releasebot:Nov 14, 2026
Seedance 2.5 4K
Runway API adds 4K output support for Seedance 2.5 text, image, and reference-mode video generation.
November 11th, 2026 - Seedance 2.5 (seedance2_5) now supports 4K output on text to video, image to video, and video to video in reference mode. Set ratio to a 4K pixel size such as 3840:2160 (16:9) or 2160:3840 (9:16). Billed at 110 credits per second of output, plus 57 credits per second of input and reference video, with the same 80 credit minimum. Draft previews remain 480p and enhanced drafts remain 1080p. See pricing, input requirements, and the text to video endpoint.
Original source - Nov 11, 2026
- Date parsed from source:Nov 11, 2026
- First seen by Releasebot:Nov 13, 2026
Seedream 5 Pro Layers billed per returned image
Runway API switches Seedream 5 Pro Layers to per-image billing, with rates based on image resolution and credits settled for images returned.
November 11th, 2026 - Seedream 5 Pro Layers (seedream5_pro_layers) is now billed per returned image instead of 18 credits per decomposition: 3 credits per image at 1K or 1.5K, and 5 credits per image at 2K or auto, counting the base image and each layer. A decomposition reserves credits for the most images it can return (17), then settles to the images returned. See pricing.
Original source - Nov 10, 2026
- Date parsed from source:Nov 10, 2026
- First seen by Releasebot:Nov 13, 2026
Seedream 5 Pro Layers on Runway Dev
Runway API adds Seedream 5 Pro Layers to split images into editable transparent layers for image-editing workflows.
November 10th, 2026 - Seedream 5 Pro Layers (seedream5_pro_layers)
Now available on Runway Dev. Split an image into editable layers for image-editing workflows.
Send one PNG or JPEG source image to the text to image endpoint, with an optional prompt and a resolution of auto, 1k, 1.5k, or 2k. Successful tasks return a base image and editable transparent PNG layers, with layer names, stacking order, and bounds when provided.
Billed at 18 credits per decomposition at any resolution; see pricing and image input requirements.
Original source Similar to Runway AI with recent updates:
- OpenAI release notes1270 release notes · Latest Nov 18, 2026
- xAI release notes313 release notes · Latest Nov 17, 2026
- Cursor release notes144 release notes · Latest Nov 6, 2026
- Eleven Labs release notes107 release notes · Latest Nov 17, 2026
- Perplexity release notes33 release notes · Latest Nov 6, 2026
- Anthropic release notes965 release notes · Latest Nov 19, 2026
- Nov 5, 2026
- Date parsed from source:Nov 5, 2026
- First seen by Releasebot:Nov 7, 2026
Kling 4.0 Flash on Runway Dev
Runway API adds Kling 4.0 Flash for generating and editing 720p videos with audio from text, images, and video references, with outputs ranging from 3 to 20 seconds.
November 5th, 2026 - Kling 4.0 Flash (kling_4_flash) is now available on Runway Dev.
Generate 3–20 seconds of 720p video (default 5 seconds) from text, a first frame, or image and video references, or edit a source video. Outputs include audio.
Use up to 10 images including the first frame, and up to 5 videos totaling at most 20 seconds including the source video.
Billed at 9.4 credits per output second without input video, or 6.3 credits per second of output plus the greater of total input-video duration and output duration, rounded up to the next credit.
See pricing and input requirements. Use the text to video, image to video, or video to video endpoint to get started.
Original source - Nov 5, 2026
- Date parsed from source:Nov 5, 2026
- First seen by Releasebot:Nov 7, 2026
GPT Image 2.6 Flare on Runway Dev
Runway API releases GPT Image 2.6 Flare on Runway Dev for prompt-based PNG generation with reference images, size, quality, and transparency options.
November 5th, 2026 - GPT Image 2.6 Flare (gpt_image_2_6_flare) is now available on Runway Dev. Generate PNG images from prompts up to 32,000 characters, with up to 16 reference images and outputCount from 1–10. Choose among 30 pixel width:height sizes across 1K, 2K, and 4K, or auto. Both quality and reasoningEffort support low, medium, and high (the default). Transparent backgrounds are supported. Billed at 1–19 credits per image, plus 1 credit per reference image per generated image. See pricing and the text to image endpoint.
Original source - Nov 5, 2026
- Date parsed from source:Nov 5, 2026
- First seen by Releasebot:Nov 7, 2026
GPT Image 2.6 Sunburst on Runway Dev
Runway API adds GPT Image 2.6 Sunburst on Runway Dev for prompt-based PNG generation with reference images, size options, and transparency.
November 5th, 2026 - GPT Image 2.6 Sunburst (gpt_image_2_6_sunburst) is now available on Runway Dev. Generate PNG images from prompts up to 32,000 characters, with up to 16 reference images and outputCount from 1–10. Choose among 30 pixel width:height sizes across 1K, 2K, and 4K, or auto. Both quality and reasoningEffort support low, medium, and high (the default). Transparent backgrounds are supported. Billed at 1–19 credits per image, plus 1 credit per reference image per generated image. See pricing and the text to image endpoint.
Original source - Nov 5, 2026
- Date parsed from source:Nov 5, 2026
- First seen by Releasebot:Nov 7, 2026
Gemini Omni 1.2 Flash on Runway Dev
Runway API adds Gemini Omni 1.2 Flash for text, image, and video generation, with outputs from 360p to 4K.
November 5th, 2026 - Gemini Omni 1.2 Flash (gemini_omni_flash_1.2) is now available on Runway Dev. Generate 3–15 seconds of video at 360p, 720p, 1080p, or 4K, with a prompt up to 4,000 characters.
Generate from text with up to 5 image references and 3 video references, animate a first frame with an optional last frame, or use video-to-video reference, extend, and edit modes.
Billed at 2, 7, 10, or 21 credits per second by resolution, plus image and input-video charges; see pricing and input requirements. Use the text to video, image to video, or video to video endpoint to get started.
Original source - Nov 1, 2026
- Date parsed from source:Nov 1, 2026
- First seen by Releasebot:Nov 13, 2026
Praxis-1: From Video Pretraining to Robot Control
Runway AI releases Praxis-1, an open-weight world action model that brings video foundation models to real-robot control. It leads open robot model benchmarks, handles everyday tasks across multiple robot types and is available with code and fine-tuning tools.
An open weight world action model that turns Runway’s video pretraining into control for real robots.
by Runway
Code, weights and integrations: see the launch post.
#1 World Action Model
Praxis-1 ranks #1 among world action models and among all open source robot foundation models on NVIDIA’s RoboLab-120 leaderboard. It also achieves a competitive success rate on real-world Franka tasks against state-of-the-art policy models.
RoboLab-120 success: world action models and open weight VLAs: Praxis-1 46.2%, FLUX 3 Action 42.9%, Atomic-WAM 39.6%, OASIS WAM 39.0%, Cosmos 3 Nano 36.8%, π0.5 (VLA) 28.0%, DreamZero 25.7%, GR00T N1.6 (VLA) 7.3%.
Everyday Tasks in Real-time
Praxis-1 controls a bimanual YAM robot and a single Franka arm through everyday tasks that need contact and coordination: placing a mug upside down on a tray, putting an item in a box and closing the lid, clearing bottles into a bin and closing a drawer, and more.
Adapts to New Robots
Praxis-1 adapts quickly to new robots. It also shows competitive results in Google RT-1 SimplerEnv against other specialized models. It can be fine-tuned on the low-cost SO-101 to solve a range of real-world tasks.
1. Introduction
Robotics deployments keep running into the same problem: real-world robot data is scarce, and collecting enough of it to train a generalist policy is expensive. Some labs respond with more teleoperation, others with more egocentric footage. We think the most scalable source is general video, the everyday footage people already record. A model trained to predict that video learns how objects behave, how hands move and what a task looks like partway through. A robot can build on that knowledge.
With Praxis-1, we turn a 27B-parameter video foundation model into an open weight world model that predicts the future in short chunks. A small action expert reads the video model’s internal state as it imagines what comes next. Each time the robot acts, Praxis-1 predicts both what the scene will look like and the motion that will get it there.
Praxis-1 fine-tuned on DROID is the top world action model and the top open-source robot foundation model on the RoboLab-120 leaderboard. On a real Franka arm it achieves a competitive success rate against the previous state-of-the-art world action model.
Praxis-1 can adapt to various embodiments, from single arms to bimanual robots and mobile manipulators. We showcase everyday tasks such as placing a mug upside down on a tray, putting an item in a box and closing the lid, clearing bottles into a bin or closing a drawer. The model can also adapt to embodiments that it hasn't seen in training data. We detail this with Google's RT-1 and SO-101 below.
2. Architecture
Praxis-1 has three parts. A video backbone predicts what the scene will look like next, and an action expert reads the predicted video features to work out the motion that produces it. Thin per-robot projections translate each robot’s poses and actions into a shared format, so one model can control different embodiments. On each call, the model takes an instruction, the latest camera frames and the robot’s current pose. It returns a predicted chunk of future video along with 1.6 seconds of actions: 24 action rows at 15 Hz. The robot executes the chunk, sends back what its cameras saw and calls again.
Figure 1 — The backbone and the action expert
Tap or hover a block for its role. Video and actions are both predicted one chunk at a time (24 action rows, 1.6 s at 15 Hz). Each robot has its own thin input and output projections; the backbone and the action expert are shared. Parts are described below.
2.1 The video backbone
Praxis-1 is built on Runway's video foundation model: a diffusion transformer split into a condition encoder and a velocity decoder (a DDT layout, after Wang et al., 2025). It works in the latent space of a frozen Wan2.2 VAE and reads instructions through a frozen Qwen3.5-2B text encoder. 17 of its 20 feed-forward layers are mixture-of-experts (MoE); only the first two encoder blocks and the first decoder block stay dense.
Each MoE layer has one shared expert that every token uses, and 128 small routed experts, of which each token picks eight. The nine experts a token passes through add up to the width of one dense layer, so the compute per token stays about the same as a dense block while the layer holds 129 experts' worth of parameters. In total the backbone has 26.9B parameters, of which about 2.8B are active for any token. Experts are balanced as in DeepSeek-V3 (DeepSeek-AI, 2024): a routing bias that needs no auxiliary loss, plus a small sequence-level auxiliary loss.
The base model was trained to generate whole clips at once, with every frame able to see every other frame, including later ones. A robot can’t see the future, so during mid-training we change the attention to be chunk-causal: the model works through time in 1.6-second chunks of six latent frames, and each chunk can only see the chunks before it. Each chunk is denoised from its own noise level while attending to itself and to clean copies of the earlier chunks. In training, those earlier chunks come from real video rather than the model’s own predictions, a setup known as teacher forcing, which lets us train on four chunks at once in a single parallel pass. At inference, the model applies the same rules one chunk at a time. By default it reads one earlier chunk of real camera frames from the robot, so it sees the same kind of clean context it saw in training. The attention mask is shown below, alongside the action expert’s.
2.2 The action expert
The expert is a dense 16-layer transformer of width 1024, with about 314M parameters. It is closest in spirit to DeVA (Zhang et al., 2026), which also keeps a separate action expert that reads several depths of the video model. Its layers alternate between cross-attention into the backbone and self-attention. Each cross-attention layer reads one of eight backbone encoder blocks, spread from early to late layers, while a shared normalization and a 3072→1024 projection bring the backbone features to the expert's width.
Per chunk the expert holds one absolute pose token, eight learned "future" tokens and six action tokens of four rows each. Its queries for chunk k see the clean backbone frames before k and the backbone's noised frames of chunk k itself. The expert therefore reads the future as the backbone imagines it. The backbone never reads the expert, but the action loss still trains the backbone through those cross-attention reads.
We also tried putting action tokens inside the backbone (a mixture-of-transformers layout). On our offline proxy, a short bimanual fine-tune scored on action error, this layout showed no clear difference from the separate expert (+0.01 cm) and cost more. Thus we kept the separate action expert and leave other architectures as future work.
2.3 Inputs and sampling
- Cameras: Exterior + wrist, each resized to a 224×416 tile, side by side on a 224×832 canvas, 15 fps
- Window: 93 frames at 15 fps in training: an anchor frame, a shorter first chunk and three full chunks
- Action row: Set per embodiment, each with its own projections. Single-arm end-effector control (DROID Franka) is 10-D: xyz, 6-D rotation, gripper; bimanual YAM is 20-D (two arms); human hands on egocentric video use 21-joint 3-D poses. Position and rotation relative to the pose at the chunk start; position and gripper normalized with 1st–99th percentile bounds and clipped
- Chunk: 24 rows = 1.6 s at 15 Hz (21 rows on the first call of a context window), executed in full
- Robot state: Absolute pose at the chunk start, z-scored; dropped out for 10% of training samples
- Per-robot parts: Action-in, pose-in and action-out projections. Backbone and expert are shared across robots
- Objective: Rectified flow on both streams. One video noise level per chunk, weighted toward high noise; one action noise level per chunk, uniform and drawn independently
- Sampler: 2 joint Euler steps over video and actions, 1 context chunk, text-only guidance on the video. How we chose the guidance settings: Finding the optimal guidance
- Size: 27.2B parameters in total, 54.4 GB in bf16; runs on one 80 GB GPU
3. Training stages
Praxis-1 learns how the world works before it learns to act. Training runs in four stages, each starting from the weights of the one before, and video prediction stays part of the objective throughout. When actions are introduced, the model learns them alongside video rather than in place of it. Each stage reaches its effective batch size through gradient accumulation on top of the GPU count.
- 01 Video pretraining backbone
- 02 Embodied-video mid-training backbone
- 03 Video & action joint mid-training backbone + expert
- 04 Embodiment-specific FT backbone + expert + embodiment heads
4. Mid-training Compute Scaling
We expect more mid-training compute to give fine-tuning a stronger base model. We tested this on two robots: Franka, which is seen in our DROID mid-training data mix, and RT-1, a new embodiment never seen in any pretraining or mid-training stage. Each point is a checkpoint from the same mid-training run, placed by the total compute spent to reach it, video pretraining included. Because later checkpoints revisit the same mid-training data, this measures more training compute rather than more data.
In our experiments on two embodiments, more mid-training did lead to better fine-tunes. We fine-tuned five checkpoints from the video and action mid-training run on DROID with the same schedule. On RoboLab-120, with the same sampler settings for every point, the first four scored progressively higher, 25.5%, 35.2%, 39.2% and 41.2%, and the fifth leveled off at 41.2%.
RT-1 is absent from the mid-training data, which allows us to test the same question on an unseen robot. We fine-tuned five mid-training checkpoints on RT-1 with the same recipe. SimplerEnv success rose from 56.3% for the earliest checkpoint to 80.0% for the latest, though not at every step (Figure 6). We see a trend of closed-loop improvement on both the seen and unseen embodiments, but the exact relationship needs further study.
5. Inference
Our model predicts actions chunk by chunk: the robot executes the whole chunk, sends back the frames it saw and calls again over a single WebSocket. Each policy call returns 24 action rows at 15 Hz, or 1.6 seconds of motion, denoised with the imagined video in two joint Euler steps. The first call of a context window returns 21 rows instead, because the first of its six latent frames covers only the current frame, not four. On a single H100 server, a call with the default chunked execution takes about 0.55–0.6 seconds round trip, about a third of the 1.6 seconds of motion it returns. By default, we run video classifier-free guidance (CFG) sequentially.
We have also explored real-time chunking (RTC): instead of waiting for a chunk to finish, the robot asks for its next chunk partway through the current one and keeps moving. The hard part is the junction. Blending the old and new plans made the arm oscillate, so the new plan instead continues from where the arm will actually be, using the model's own motion. On our bimanual station, this cut the time spent holding still about fivefold (from 32% to 5% in a session where each call took 0.8–0.9 seconds, and from 16% to 3% with a faster server at about 0.3 seconds per call). On a five-object pick-and-place into a bag, RTC reached a comparable success rate (24/30 vs 28/30 chunked) and finished faster (median 9.1 vs 15.2 seconds to success). RTC is opt-in in our release repo; chunked execution is the default.
Released server latency
- Bimanual YAM, chunked (default): 545 ms
- Bimanual YAM, RTC + streaming optimization (opt-in): 128 ms
- Single-arm Franka (DROID), chunked (default), video CFG 4: 522 ms
- Single-arm Franka (DROID), RTC + streaming optimization (opt-in), video CFG 4: 124 ms
- Motion returned per call: 24 rows, 1.6 s at 15 Hz (21 rows on the first call)
6. Results
Every result in this section comes from the same Praxis-1 checkpoint, fine-tuned on DROID using the recipe in Training Stages. We evaluate it in simulation on RoboLab-120, and on a real Franka FR3 arm operated by Positronic.
On NVIDIA’s RoboLab-120 leaderboard, as of publication. RoboLab-120 is NVIDIA's simulation benchmark of 120 tabletop tasks for a DROID-style arm. Under the official protocol of 10 trials per task, Praxis-1 finetuned on DROID solved 555 of 1,200 episodes (46.2%), the highest of any world action model and of any open source robot foundation model on the leaderboard. FLUX 3 Action follows at 42.9%. Praxis-1 is also ahead of every open weight VLA on the board, including π0.5 and GR00T N1.6.
Four entries score above Praxis-1 on RoboLab-120. The top two are agents that drive a frontier model directly rather than trained policies: AMP1 (86.0%), listed as a zero-shot frontier-model agent, and SenseNova-RoboRSI (57.1%), which runs a frontier vision-language model as the policy. SenseNova-RoboRSI is far slower than Praxis-1: its authors report about 55 minutes of computation per episode for 8–76 seconds of commanded motion on a comparable tabletop benchmark (RoboDojo), and call the execution time impractical for deployment (arXiv 2610.22471). Praxis-1 computes each 1.6-second chunk of motion in about 0.55–0.6 seconds on one H100. The other two, Lakeside (51.0%) and BiMind v0.5 (49.2%), are closed models trained partly on simulation data, and BiMind v0.5 was fine-tuned on RoboLab data.
7. Everyday Tasks in Real-time
Topping a leaderboard is one thing; running on a real robot is another. Here Praxis-1 runs two real robots on everyday tasks: a bimanual YAM robot, whose two arms work together on tasks one arm cannot do alone, and a single Franka arm. Every clip below shows the policy in control, with no teleoperation, played at 1x.
The YAM fine-tunes follow the same recipe as DROID: they start from an action mid-training checkpoint and learn from roughly 1,400 to 2,600 hours of bimanual YAM recordings with three cameras.
Collected on bimanual YAM stations at the Runway office, real speed; each clip is labeled with its instruction.
- “Pick up the purple comb on the right then put the purple comb into the bag” 1x speed.
- “Pick up the red die, black pen and yellow sponge and place each one in an empty plate” 1x speed.
- “Place the white mug upside down on the yellow tray” 1x speed.
- “Pick the teddy bear and place it on the towel” 1x speed.
- “Pick up the lid and close the box” 1x speed.
- “Pick up the red pepper and place it in the mug” 1x speed.
- “Pick up the cutting board and place it in the yellow tray” 1x speed.
- “Pick up the plastic bottles and place them in the silver garbage” 1x speed.
Collected by Positronic on a Franka FR3 station, real speed; each clip is labeled with its instruction.
- “Put the spoon in the mug” 1x speed.
- “Close the drawer” 1x speed.
- “Move the cup to the right” 1x speed.
- “Put the pen in the cup” 1x speed.
- “Put the marker on the table” 1x speed.
- “Put the fork on the plate” 1x speed.
- “Put the yellow block in the blue cup” 1x speed.
Frontier multimodal models can now drive robot arms zero-shot, through tool or code interfaces. GPT-6 Astra placed above every public entry on RoboDojo when it was first evaluated, with 22.5% success (Zhang et al., 2026), and reaches 76.0% on RoboTwin 2.0 (Zhou et al., 2026). These agents are general: they solve a wide range of tasks zero-shot, and today they lead the RoboLab-120 leaderboard, where the top two entries are LLM agents (86.0% and 57.1% success). Their drawbacks are latency and cost. On the AA-Robotics benchmark, Astra averaged 30 minutes and $16.36 per tabletop task, against 20 seconds for a person (Artificial Analysis, 2026), and Zhou et al. report that its inference latency hinders practical deployment.
To measure this gap ourselves, we ran Praxis-1 and GPT-6 Astra on the same YAM teleoperation station, through the same console, on the same three tasks. On average, Astra took 15.6x longer to finish a task and cost 51x more per completed task (Figures 11 and 12).
Time and cost per task, Praxis-1 vs GPT-6 Astra
- Pen cap: Praxis-1 0:15; Astra 7:48; faster with Praxis-1 31x; Praxis-1 $0.08; Astra $3.47; cheaper with Praxis-1 43x
- Sponge in bag: Praxis-1 0:22; Astra 2:46; faster with Praxis-1 7.4x; Praxis-1 $0.03; Astra $1.20; cheaper with Praxis-1 49x
- Die in bag: Praxis-1 0:11; Astra 2:04; faster with Praxis-1 11.5x; Praxis-1 $0.01; Astra $0.93; cheaper with Praxis-1 79x
- All three: Praxis-1 0:16; Astra 4:13; faster with Praxis-1 15.6x; Praxis-1 $0.04; Astra $1.87; cheaper with Praxis-1 51x
Frontier LLM agents are advancing quickly, and the two approaches are complementary: an agent can plan a new task, and a fast policy like Praxis-1 can carry out each step in real time. We are studying how to pair them as future work (see Where we go from here).
8. Adapts to New Robots
8.1 Google RT-1
To further test how Praxis-1 approaches a robot it hasn't seen before, we fine-tuned on RT-1, Google's mobile manipulator, whose recordings never appeared in our mid-training mix. From the latest mid-training checkpoint, we fine-tuned for 30k steps (about 28 hours on 32 H100s) on the full RT-1 train split: about 78k episodes and 95 hours, head camera only. We scored the model on SimplerEnv's official visual-matching grid (864 episodes over four tasks), the benchmark most RT-1 policies report.
Praxis-1 fine-tuned on RT-1 reaches 80.0% averaged over the four tasks, above RT-1 itself (52.4%). Comparatively, FLUX 3 Action reaches 39.7% when we fine-tune it on the same RT-1 data from its base checkpoint with Black Forest Labs’ published DROID fine-tuning recipe.
8.2 SO-101
SO-101 is a low-cost, open-source single arm built for hobbyists and researchers, and its community has uploaded thousands of small LeRobot datasets. We gathered 3,956 public SO-100/SO-101 datasets from 1,783 contributors, about 1,260 hours, mostly recorded at 30 fps and 640x480. Each has one to three scene cameras, and 2,438 of them (about 860 hours) also have a wrist camera.
Camera names in community datasets are unreliable: a camera called “phone” or “top” is often mounted on the wrist. So we labelled every camera from its video, and counted it as a wrist camera only when two checks agreed: motion (a fixed scene camera keeps most of its pixels still over 64 frames, a wrist camera does not) and appearance (a zero-shot CLIP model).
We held out 102 whole datasets (64 hours) from fine-tuning for validation, and fine-tuned Praxis-1 on the training episodes of the rest, about 1,070 hours from 3,814 datasets. Each sample uses one scene camera, plus the wrist camera when there is one; wrist datasets are also sampled with the scene camera alone, so one model runs with or without a wrist camera. The model predicts the arm’s five joints (as offsets from the current state) and the gripper at 15 Hz, in 1.6-second chunks. Instructions are the clip’s task or auto-captioned sub-steps that fall inside each training window. Each clip below shows the policy in control on a single instruction, played at 1x.
- “Fold the towel in half” 1x speed.
- “Grab the sponge then wipe the green plate with it” 1x speed.
- “Put the pink eraser in the purple bowl then drop it” 1x speed.
- “Put the teddy bear on the towel” 1x speed.
9. Finding the optimal guidance
Praxis-1 samples each action chunk with two denoising steps and applies classifier-free guidance to the instruction only: the unguided pass keeps the video history and the pose but swaps the instruction for a blank caption. The video guidance scale sets how strongly the predicted video follows the instruction, and a scale of 1 turns it off. The second knob is the flow shift, which warps the time grid of the two denoising steps. A larger shift leaves more of the denoising to the second step. Video and actions share one shift, which we tune for both together; giving each its own shift is left for future work.
We set both knobs with a step-wise search on a fixed set of simulated environments, starting from no video guidance at shift 8. Adding video guidance helped, and raising it together with the shift helped more, ending 4.0 points above the start; pushing either knob further scored lower. We ship video CFG 4 with shift 12.5 and two denoising steps. These guidance settings are embodiment-specific: they were tuned for the DROID Franka, and other robots need their own.
10. Offline evaluation
Alongside closed-loop benchmarks, we judge the model open loop in two ways: how accurate its actions are, and whether the future it imagines follows the instruction.
10.1 Does the error come from the predicted video or the predicted actions?
We measure action error as the distance between the predicted and ground-truth absolute end-effector positions, on 32 held-out DROID windows. We use two conditioning modes to separate two sources of error: imagining the wrong future, and reading the wrong motion off the right one.
- Deployment: the model samples its own video and actions together, as it does on the robot.
- Oracle: we give the model the true future video, noised to the sampler's level, and sample only the actions, following the oracle test in mimic-video (Pai et al., 2025).
We observe that most of the remaining error comes from which future the video model imagines, not from the action expert, so improvements to future video prediction should pay off most. Given the real future, the error is about one-fifth as large. Note that deployment error is not a fully reliable metric: it combines future-video and future-action prediction, and several different futures can accomplish the same task. We discuss a better way to evaluate generated videos in the next section.
10.2 Can the model picture the task before doing it?
To understand how the model plans and imagines the future, we give it a single first frame of a 3D-built scene instead of a real episode (only the background comes from a real image, with objects of various colors, sizes and positions added), the instruction and the robot’s real starting pose. It then rolls itself out for 12.6 seconds, predicting video and actions together with no ground truth after the first frame. A vision-language judge watches each predicted video and counts it as following the instruction only if the right object is acted on, it goes to the right place and the action is correct. The scenes are built with look-alike objects, to test how robustly the model follows the prompt. Instruction following in imagination is a proxy for how well the model understands the task; how it relates to real-world success still needs further study.
Examples from the DROID prompt set: 12.6-second open-loop self-rollouts of Praxis-1 finetuned on DROID, starting from the first frame shown, with the judge’s verdict.
The same imagined-rollout method also evaluates the mid-training base model, which is not fine-tuned on any specific embodiment. These long-horizon examples come from that base model on a bimanual YAM station, shown for video quality rather than task success. Each starts from a single first frame of a 3D-built scene instead of a real episode, and the model imagines all three cameras for up to 20 seconds with no robot and no actions. We also change the instruction during each rollout to see if the model can react to it in its video prediction, as listed above each example.
10.3 Can the model predict whether a task will succeed?
We want to understand if the model’s prediction of its own future also signals whether the task will succeed. In internal research on an earlier, unreleased version of the model fine-tuned for the Franka arm, we froze its weights and trained a small head on its representation of the next predicted chunk to estimate the chance that the episode succeeds from that point. Its only supervision was whole-episode outcomes, with no hand grading. On 272 real Franka episodes from sessions and policies it never saw, the score separated successes from failures after four chunks (about 6 seconds) with an AUROC of 0.77 (95% CI 0.70–0.83); elapsed time alone gets 0.64. Averaged per policy, it also ranked nine checkpoint and sampler combinations by their real success rate with a Spearman correlation of 0.79, though the interval is wide (0.19–1.00).
This is early evidence. The head doesn’t yet read the instruction, so it gives a clean pick-and-place the same score whether or not it was the right task. Still, the result points to the direction we care about most, which is a model that learns from its own video predictions by using its imagined future to score and then improve its actions. We are actively researching this direction toward a truly self-improving world action model.
11. Quantization
Praxis-1 also holds up with its routed MoE experts rounded to 4-bit NVFP4. We fine-tuned the released checkpoint for 5k more steps with quantization-aware training, rounding the routed expert matrix multiplies to NVFP4 (E2M1 values in blocks of 16, E4M3 block scales and an FP32 global scale). We sampled 600 episodes from RoboLab-120 and ran them head-to-head on the two checkpoints: the NVFP4 model succeeds in 279/600 (46.5%) and the same fine-tuned checkpoint run in bf16, with the rounding switched off, in 287/600 (47.8%). The two are close. We repeated the test on the bimanual YAM, with a 5k-step quantization-aware fine-tune of a bimanual YAM checkpoint on a real station: on a five-object pick-and-place into a bag, the NVFP4 model succeeded in 26 of 30 rollouts (87%) against 28 of 30 (93%) for the bf16 checkpoint it was fine-tuned from, with grab quality 2.73 vs 2.40 (out of 3) and mean time to success 20.7 vs 20.9 seconds. None of these differences is significant at 30 rollouts per arm. We will release the NVFP4 checkpoints in our Hugging Face repository. Packing the routed experts in NVFP4 shrinks the policy weights from 54.4 GB to 17.5 GB.
12. Limitations of Praxis-1
- Look-alike objects. The most common real-robot failure is reaching for an object that resembles the target, and the wrong choice is usually already visible in the imagined video. This is consistent with Section 10.1, where most action error comes from the imagined future.
- Vague instructions. The policy does better when the instruction names the object’s attribute (color, size); vague instructions are noticeably harder. When the instruction alone does not single out the target, the model has to guess among the plausible objects in the scene, and in a scene with look-alike objects it can pick the wrong one.
- Long tasks. Multi-step tasks work when each step gets its own instruction, but chaining several steps from a single prompt remains hard. The model has to keep track of which steps are already done and move on to the next one without being told. Pairing it with a language or vision-language model that breaks a task into steps is one of the directions in Section 13.
13. Where we go from here
We are releasing three checkpoints on Hugging Face under the Apache 2.0 license: Praxis-1 fine-tuned on DROID, Praxis-1 fine-tuned on the bimanual YAM and the base mid-trained model. They come with the model code in our GitHub repository, a policy server, a MuJoCo simulation of the same setup and code for both full and LoRA fine-tuning. See more details in our launch post.
We’ll continuously invest in world action model (WAM) research, especially focusing on long-term reasoning, agentic planning and self-improvement loops.
We are excited to deploy our policy with partners on real tasks. If you’d like to work with us on this, get in touch.
Original source - Oct 28, 2026
- Date parsed from source:Oct 28, 2026
- First seen by Releasebot:Nov 2, 2026
Ideogram Vector on Runway Dev
Runway API adds Ideogram Vector, enabling prompt-based SVG illustration generation with reference images and optional seeds, plus image-to-vector conversion with adjustable shape complexity.
October 28th, 2026
Ideogram Vector (ideogram_vector) is now available on Runway Dev. Generate one scalable SVG illustration from a prompt of up to 9,500 characters, guided by up to 5 reference images, with an optional seed. quality is low, medium (the default), or high, billed at 14, 17, and 22 credits per SVG, or 35 credits at high with reference images.
To convert a source image into vector artwork, send one imageUri to POST /v1/image_to_vector, with vectorizationMode set to default (the default) or simple for fewer shapes; conversion is billed at 11 credits per SVG. Use the text to image endpoint to get started.
Original source - Oct 27, 2026
- Date parsed from source:Oct 27, 2026
- First seen by Releasebot:Nov 2, 2026
Latest updates
Runway AI launches Runway Ads, which analyzes top-performing Meta, Google, and TikTok ads, creates on-brand variants, and publishes approved versions.
All Plans
Runway Ads
Runway Ads finds your next winning ad for you. It reads your best performers on Meta, Google and TikTok, generates on-brand variants for every placement and publishes the ones you approve. Included with every paid plan.
Original source - Oct 27, 2026
- Date parsed from source:Oct 27, 2026
- First seen by Releasebot:Oct 27, 2026
Seedance 2.0 bitrate
Runway API adds standard and high bitrate options for Seedance 2.0 video generation, with high bitrate preserving more detail at the same rate.
October 27th, 2026 - Seedance 2.0 (seedance2), Seedance 2.0 Fast (seedance2_fast), and Seedance 2.0 Mini (seedance2_mini) now accept bitrate on text to video, image to video, and video to video. bitrate defaults to standard, which is the previous output. Set bitrate to high to keep more detail with fewer compression artifacts, at a larger file size. Both settings are billed at the same rates. Use the text to video, image to video, and video to video endpoints to get started.
Original source - Oct 26, 2026
- Date parsed from source:Oct 26, 2026
- First seen by Releasebot:Oct 27, 2026
Workflows guides
Runway API adds workflows for building node-graph pipelines, publishing versioned endpoints, and running every step with one request.
October 26th, 2026 - Workflows now have guides on Runway Dev. Build a multi-step pipeline as a node graph in the Runway app, publish it as a versioned endpoint, and run every step with one request to POST /v1/workflows/{id}. See Workflows to get started.
Original source - Oct 23, 2026
- Date parsed from source:Oct 23, 2026
- First seen by Releasebot:Oct 27, 2026
Latest updates
Runway AI brings App Studio to paid plans, letting users build and host creative apps with its models, their brand, and files.
Paid Plans
App Studio
Describe the creative app you need in plain language, and App Studio builds it with Runway’s models, your brand and your files, then hosts it for you.
Original source
Curated by the Releasebot team
Releasebot is an aggregator of official release notes from hundreds of software vendors and thousands of sources.
Our editorial process involves the manual review and audit of release notes procured with the help of automated systems.