MiniMax responded to SGLang including MiniMax H3 in the TPU support list of SGLang-TorchTPU, expressing thanks and saying it looks forward to more open-source work around efficient video inference. SGLang said that SGLang-TorchTPU was jointly built by SGLang, the Google TPU team, and the Meta PyTorch team, has been used for Meta's production inference, and supports models including Muse Spark Open, GLM-5.3 Flash, Kimi K3, Ling 3.1 Flash, and MiniMax H3. There are currently no facts about subsequent developments.
fal launched Kling 4.0 on October 21, 2026. According to information released by fal, this is a new-generation native multimodal video model that improves the stability of complex motion, reference consistency for characters and scenes, and control precision for cinematic storytelling. Currently, the model is available only through fal UX and fal Agent, with API access coming soon. Apart from this launch fact, there are no other subsequent developments with qualified factual evidence.
Ethan Mollick said he had Opus 5.5 generate a video using the prompt he uses when testing all AI image generation models, "an otter using wifi on a plane." The account appears in his article reviewing the exponential development of AI over the past 30 months. During the same period, on the day his new book Co-Existence was released, Mollick used charts to review the AI progress in the 30 months since the release of his previous book Co-Intelligence, and said the charts in that article were completed by Opus 5.5 in Claude Projects orchestrating multiple agents. Currently there is evidence showing that this video generation action occurred in the above-mentioned review context, with no subsequent developments yet.
Stanford course CME 295 Transformers & Large Language Models has been updated to Lecture 4, with both the video and lecture notes now publicly available. The topic of this lecture is Reinforcement learning with LLMs, covering everything from reinforcement learning and reward design to PPO, GRPO, RLVR, and DAPO, explaining each of the notations one by one. The course is taught by Afshine and Shervine Amidi, with 9 lectures in total, covering the Transformer architecture, attention variants, RoPE, LoRA fine-tuning, RLHF/GRPO training, Flash Attention inference optimization, MCP and AI Agents, and diffusion language models. Currently, Lecture 4 is public, and no evidence of updates for subsequent lectures has been seen yet.
Stanford course CME 295 Transformers & Large Language Models has been updated to Lecture 4, with both the video and lecture notes now publicly available. The topic of this lecture is Reinforcement learning with LLMs, covering everything from reinforcement learning and reward design to PPO, GRPO, RLVR, and DAPO, explaining each of the notations one by one. The course is taught by Afshine and Shervine Amidi, with 9 lectures in total, covering the Transformer architecture, attention variants, RoPE,
On October 20, 2026, Wan announced that Wan3.0 Prime ranked first in the New & Trending section of the OpenRouter video leaderboard, while Wan3.0 ranked third. Wan also thanked creators for choosing Wan. This news was released by Wan as the first party, and there are currently no subsequent developments.
Gemini Notebook launches a collaborative creation feature for slideshows and video overviews. Users can first select a focus topic, then set the format, length, language, and visual style, then review and edit the slide outline or video production notes, and finally generate the content. This feature is available to all Pro and Ultra subscribers. There are currently no subsequent developments.
Gemini Notebook launches a collaborative creation feature for slideshows and video overviews. Users can first select a focus topic, then set the format, length, language, and visual style, then review and edit the slide outline or video production notes, and finally generate the content. This feature is available to all Pro and Ultra subscribers. There are currently no subsequent developments.
Ethan Mollick said he had Opus 5.5 generate a video using the prompt he uses when testing all AI image generation models, "an otter using wifi on a plane." The account appears in his article reviewing the exponential development of AI over the past 30 months. During the same period, on the day his new book Co-Existence was released, Mollick used charts to review the AI progress in the 30 months since the release of his previous book Co-Intelligence, and said the charts in that article were completed by Opus 5.5 in Claude Projects orchestrating multiple agents. Currently there is evidence showing that this video generation action occurred in the above-mentioned review context, with no subsequent developments yet.
Ethan Mollick said he had Opus 5.5 generate a video using the prompt he uses when testing all AI image generation models, "an otter using wifi on a plane." The account appears in his article reviewing the exponential development of AI over the past 30 months. During the same period, on the day his new book Co-Existence was released, Mollick used charts to review the AI progress in the 30 months since the release of his previous book Co-Intelligence, and said the charts in that article were compl
A Tsinghua University team has proposed PPMem, a memory method for video world models. According to a report on October 20, 2026, the method replaces the volumetric 3D memory of the earlier SPMem with planar primitives to store explored scenes for video world models. Each planar primitive is parameterized by only 7 numerical parameters, including a center depth, two rectangular extents, and a rotation quaternion, making it several orders of magnitude more compact than per-pixel representations. When the camera returns to a previously visited location, the stored planes are projected back into the field of view to guide the generation of new frames, thereby reducing the spatial information that must be retained and processed and preventing familiar objects from changing or disappearing upon revisiting. The work is currently presented externally under the name PPMem, and no further progress has been disclosed.
A Tsinghua University team has proposed PPMem, a memory method for video world models. According to a report on October 20, 2026, the method replaces the volumetric 3D memory of the earlier SPMem with planar primitives to store explored scenes for video world models. Each planar primitive is parameterized by only 7 numerical parameters, including a center depth, two rectangular extents, and a rotation quaternion, making it several orders of magnitude more compact than per-pixel representations.
A Tsinghua University team has proposed PPMem, a memory method for video world models. According to a report on October 20, 2026, the method replaces the volumetric 3D memory of the earlier SPMem with planar primitives to store explored scenes for video world models. Each planar primitive is parameterized by only 7 numerical parameters, including a center depth, two rectangular extents, and a rotation quaternion, making it several orders of magnitude more compact than per-pixel representations. When the camera returns to a previously visited location, the stored planes are projected back into the field of view to guide the generation of new frames, thereby reducing the spatial information that must be retained and processed and preventing familiar objects from changing or disappearing upon revisiting. The work is currently presented externally under the name PPMem, and no further progress has been disclosed.
Stanford course CME 295 Transformers & Large Language Models has been updated to Lecture 4, with both the video and lecture notes now publicly available. The topic of this lecture is Reinforcement learning with LLMs, covering everything from reinforcement learning and reward design to PPO, GRPO, RLVR, and DAPO, explaining each of the notations one by one. The course is taught by Afshine and Shervine Amidi, with 9 lectures in total, covering the Transformer architecture, attention variants, RoPE, LoRA fine-tuning, RLHF/GRPO training, Flash Attention inference optimization, MCP and AI Agents, and diffusion language models. Currently, Lecture 4 is public, and no evidence of updates for subsequent lectures has been seen yet.
fal says FLUX 3 Max achieved 1,139 Elo on a third-party benchmark, entering the top tier of frontier image-to-video models based on independent human preference evaluations. The model was launched through a collaboration between fal Research and bfl_ai. Earlier reports said that through dedicated post-training and inference optimization, the model tied for first place in image-to-video quality on an independent third-party benchmark, generating a 10-second 720/768p video in about 12 seconds; other reports said it scored 1,139 on the Artificial Analysis image-to-video index, with generation speeds up to 22 times faster than Seedance 2.5. fal Research says it is continuing to post-train the model and will subsequently improve quality and performance.
On October 20, 2026, Wan announced that Wan3.0 Prime ranked first in the New & Trending section of the OpenRouter video leaderboard, while Wan3.0 ranked third. Wan also thanked creators for choosing Wan. This news was released by Wan as the first party, and there are currently no subsequent developments.
On October 15, 2026, Kling AI announced the full availability of KLING 4.0, which users can access through a redesigned creation interface and which supports integration into video creation workflows via MCP and CLI; the company also previewed upcoming 4K and 1080p 10-bit HDR video output, as well as API access for enterprise customers and developers. On October 17, fal previewed the imminent launch of Kling 4.0, saying it can generate videos up to 30 seconds long in a single pass, supports native audio, and can use up to 15 reference materials to guide characters, scenes, and camera movement. On October 20, according to The Decoder, Kling AI released version 4.0, which can generate videos up to 30 seconds long in a single pass, supports input of up to 10 keyframe images, extends prompt understanding to 8,000 tokens, adds dual-channel stereo audio and lip sync in more than ten languages, and introduces Kling 4.0 Flash, a lightweight version supporting up to 720p and 20 seconds. On October 22, Kling announced the official opening of the KLING 4.0 API, allowing developers to integrate it into their own products and workflows, but did not disclose details such as interface capabilities, pricing, or availability scope.