OpenAI launches GPT-Live as Alibaba debuts Qwen3.8-Max

Alibaba introduced Qwen3.8-Max on August 2, calling it its largest and most capable Qwen model to date.¹ The multimodal model uses a mixture-of-experts architecture that activates only part of its parameters for each request.¹² It is already accessible through Alibaba Cloud Model Studio APIs and QwenWork, including to developers outside China.² Alibaba plans to release the model weights on Hugging Face and ModelScope the following week.¹

Alibaba is positioning the model for coding and workplace tasks as it competes with other major AI providers.¹³ Its API rates may appeal to applications with substantial token usage, though actual costs depend on the mix of input and output.⁴ Published performance comparisons for agentic computer use rely on company claims and should not be treated as independently confirmed results.⁴ Organizations seeking to run the model themselves must wait for the planned weights release.¹

Meta introduced Muse Code as a beta terminal coding agent on August 5, 2026. It runs on the newly released Muse Spark 1.2 model and is designed to plan, implement, and validate changes across large code repositories. For complex tasks, it can use persistent background agents that work asynchronously. Meta also expanded global access to Muse Spark 1.2 through the Meta Model API.¹

The release gives Meta a tool for software engineering tasks that extend beyond individual code suggestions. Meta says Muse Spark 1.2 performs better at coding than its predecessor, but its announcement does not include a detailed benchmark table for that claim. Teams should test performance on their own repositories and check whether their chosen pricing tier permits prompts and completions to be used for training.¹²

On August 3, OpenAI detailed the technology behind GPT-Live for ChatGPT Voice.¹ The system can listen and speak simultaneously, while more complex reasoning and tool use remain asynchronous.¹² It handles interruptions during a conversation rather than relying on a turn detector for each exchange.¹ GPT-Live-1 is becoming the default for paid users, with GPT-Live-1 mini for Free users. The rollout covers ChatGPT on iOS, Android, and the web in supported regions.²

The architecture replaces the listen-then-respond pattern of Advanced Voice Mode with a more continuous conversation.¹² Spoken responses also appear alongside streamed text.¹² That could matter when people interrupt an assistant or ask follow-up questions. The new Live experience did not support video or screen sharing at launch.² Access expanded to Business workspaces on August 7, with charges for usage beyond the included allowance.³

Google has expanded Gemini Spark internationally and is connecting its web-browsing capabilities to Chrome. With a user’s permission, Spark can use signed-in accounts and saved passwords to complete tasks on the web. The Chrome feature is initially rolling out in the U.S. to Google AI Pro and Ultra subscribers, with more regions planned.¹²

The integration lets the assistant handle tasks that require a login, such as scheduling apartment viewings or researching flight options. Google says it hands sensitive steps, particularly payments, back to the user for confirmation and includes protections against prompt injection. For businesses, granting access to accounts is therefore a significant security decision. Spark’s broader international expansion does not mean its Chrome feature is available in every market yet.¹³

Mistral AI released Shieldstral on August 4, 2026, as an open-weights model with 3 billion parameters for moderating text and images.¹² Mistral says it can run on a single NVIDIA GPU with 16 GB of memory.¹² Users can provide moderation rules in plain language with each request. Shieldstral then scores content against those rules without retraining the model.²

The approach could let teams apply different content policies with the same locally deployed model rather than adapting a model for every policy change.² Local deployment may also matter when sensitive content should not be sent to an external moderation service. Companies should test performance on their own policies: the published benchmark results come from Mistral and do not establish how well the model will handle every use case.³²

Cloudflare released an early preview of @cloudflare/computer on August 3, 2026.¹ The open-source runtime is designed to give AI agents their own working environment with a persistent file system.¹² It uses lightweight isolates or full Linux containers depending on the task.¹² Built-in tools let agents read, write, and edit files.¹²

Cloudflare is targeting applications that need agents to retain files and working state across execution steps.¹² The file system connects the different execution environments.² The software remains a preview, and the product announcement does not specify pricing or performance figures.¹² Cloudflare OS, a separately announced browser-based workspace, is a different project rather than the same runtime.³

MiniMax introduced H3 on July 31, 2026, as a model that handles text, images, video, and audio together and generates video with native stereo sound. Clips can run up to 15 seconds, with a stated maximum resolution of 2K.¹ On August 3, the company announced it was opening the model. Its initial announcement said the weights would become available in the following days.¹²

The official repository now lists two task-specific checkpoints and a separate path for 2K regeneration. That distinction means the maximum resolution should not be read as a promise for every direct generation run.³ For companies, H3 adds another option for generating video and sound within one model. Before deploying it locally, they should check which weights are available and what license terms apply.

Liquid AI released LFM2.5-2.6B on August 4, 2026. The 2.6-billion-parameter model is designed for agents that carry out multistep tasks and use tools. Base and post-trained versions are available with open weights on Hugging Face. Liquid AI says the model can run entirely on phones, laptops, PCs, and robots without sending requests to a cloud service.¹²

The release emphasizes local execution rather than the size of cloud-based models. That could matter to companies that want data to stay on a device or need applications to work without a network connection. Liquid AI reports a memory footprint below 2.5 GB, but its performance figures do not guarantee the same results on every device or task. The weights use the LFM Open License, whose terms teams should check before commercial deployment.¹³

The paper “Training nGPT” describes a training method for nGPT. In a model with 14 billion total parameters, it reached the same validation loss using approximately half as many training tokens as an unnormalized baseline trained with AdamW.¹ The paper reports a research result, not a product launch or an available service.¹

The method changes aspects of optimization and how parameter updates are controlled.² The comparison shows lower token requirements for the tested model at the same validation loss. It does not establish that other architectures or training methods would see the same reduction.¹ That distinction matters for companies training large models: fewer tokens could improve training efficiency, but the reported comparison does not establish lower total costs or better performance on downstream tasks.¹

Cursor released the source code for Mixture-of-Kittens (MoK) on August 4, 2026. The training kernel targets mixture-of-experts models on NVIDIA NVL72 systems.¹² According to Cursor, MoK combines communication and computation in a single deterministic kernel.¹² In layer-level tests, it ran up to 2.37 times faster than the strongest public baselines. Cursor also reports 1.41 times the training throughput of its previous DeepEP-based system.¹²

The code is primarily relevant to teams training large MoE models on compatible NVIDIA hardware.² For those teams, the key question is whether communication and computation create similar bottlenecks in their own training runs. The 2.37× figure is a peak result from layer-level tests, not a general speedup for end-to-end training.¹² The published performance figures come from Cursor and do not establish independent verification.¹²