OpenAI launches GPT-5.6 and ChatGPT Work as Moonshot AI unveils Kimi K3

OpenAI released GPT-5.6 in early July 2026 alongside ChatGPT Work, an agent for creating documents, presentations, and websites.¹² Reuters reported that ChatGPT Work runs on GPT-5.6.² Its web and mobile rollout began July 9 for Pro, Enterprise, and Edu users, with access for Plus and Business expected over the following days.²

OpenAI also positioned GPT-5.6 as the preferred model in Microsoft 365 Copilot.³ This places the model in both a new agent product and existing office software. ChatGPT Work is not available on every platform and subscription at the same time, however: access is rolling out in stages.²⁴ For companies, the practical question is which features are available in the products their teams already use.

Moonshot AI introduced Kimi K3 on July 16. The company says the multimodal model has 2.8 trillion parameters and supports contexts of up to one million tokens. It is designed for coding, knowledge work, and agentic tasks. K3 is available through Kimi’s website, app, and API; Moonshot says it will release the full model weights on July 27.¹²

Until that planned release, users cannot run the weights themselves.¹ Moonshot attributes faster decoding over long contexts to Kimi Delta Attention and Attention Residuals.¹ The company also reports strong results against proprietary models in selected tests, but those comparisons have not been independently confirmed.³ For businesses, the model’s value will depend not just on its capabilities but also on costs and latency in their own workloads.

Thinking Machines Lab introduced Inkling on July 15, 2026, calling it its first model trained from scratch and its first open-weight release.¹ Its full weights are available to download, and developers can fine-tune it on Tinker.¹ The mixture-of-experts model has 975 billion total parameters, with 41 billion active parameters, and natively handles text, images, and audio.¹ Reuters reported that Inkling is also available on other developer platforms.²

The company positions Inkling as a foundation for customization rather than a finished chatbot.¹ That makes the release relevant to teams looking to further train a model for their own applications. Reuters described it as one of the few US alternatives to large open-weight models from Chinese AI labs.² Thinking Machines directs users to its documentation for Tinker pricing; the announcement does not include a price list.¹

On July 14, 2026, PrismML released Bonsai 27B, a compressed multimodal version of Qwen3.6-27B. The company says its 1-bit binary build is about 3.9 GB and can run locally on an iPhone 15 or newer. A 1.58-bit ternary build is also available.¹² PrismML CEO Babak Hassibi told CNBC that Apple and other companies were evaluating the technology. Apple did not comment.²

Bonsai 27B compresses an existing model rather than introducing a newly pretrained, 27-billion-parameter foundation model.¹³ That makes the release relevant to companies considering on-device inference instead of relying entirely on servers. Phone compatibility and reported speeds are PrismML claims. Teams still need to test whether performance and output quality meet their requirements on their own hardware.¹⁴

In July 2026, OpenAI presented a claimed proof of the Cycle Double Cover Conjecture and attributed it to GPT-5.6 Sol Ultra. A proof document and the prompt used to produce it were shared publicly.¹² A manuscript describing the claimed solution is also available on arXiv.³ Publication alone does not establish that the argument is complete and correct.

The Cycle Double Cover Conjecture has remained open in graph theory for decades.⁴ The case illustrates how AI systems are being applied to difficult proof tasks. Reports say multiple subagents worked in parallel.¹ The next step is scrutiny by mathematicians: a public manuscript is not a substitute for independent review. Until then, this remains OpenAI’s claim, not a confirmed mathematical result.³⁴

Ant Group’s InclusionAI team described Ring-2.5-1T-Zero in research presented in July. The trillion-parameter reasoning model uses the Ling architecture. According to the report, the team trained it with reinforcement learning and verifiable rewards, without human-annotated training data. Broad public availability of this model has not been established.¹

The work examines whether reasoning capabilities can develop at this scale without human annotations. The team reports behaviors including self-verification and parallel approaches to solving problems during training. Those observations and the reported math benchmark results have not been independently verified. Ant Group also published technical information on the related Ring 2.6 model generation around the same time; that publication does not establish public access to Ring-2.5-1T-Zero.¹²

The independent inference project Colibrì enables Z.ai's GLM-5.2 to run on a PC without a GPU. It keeps shared model components in RAM and loads needed experts from an NVMe SSD.¹ Z.ai had already released GLM-5.2 as an open-weight model in June; Colibrì is not a new model version from the company.²¹

GLM-5.2 is a mixture-of-experts model with about 744 billion parameters, only a portion of which are used for each output.¹ Loading experts from storage reduces the amount of RAM needed, but does not establish how fast the setup will be on specific tasks.¹ GLM-5.2 is also being optimized for GPU-based systems in other deployments.³ The CPU approach offers another way to test the model locally, rather than a demonstrated replacement for GPU serving.

Unsloth released quantized Qwen3.6 variants in July, including NVFP4 versions of the 27B and 35B-A3B models for local inference.¹² The models are available through Hugging Face, and Unsloth said it was adding Qwen3.6 support across its software stack.²³ The company advertises faster inference with the new quants on supported hardware.¹²

On July 20, AMD officially confirmed Unsloth support for its hardware, including Radeon RX 7000 and 9000 GPUs. That support also covers GGUF and llama.cpp inference paths; according to AMD, native AMD inference for Qwen3.6 is enabled by default.⁴ This gives users of AMD systems a path to running the models locally. However, the published performance measurements for Unsloth’s new quants use NVIDIA hardware and do not establish a matching speed gain on Radeon.⁵²

Robbyant, the robotics arm associated with Ant Group, open-sourced LingBot-Map in April 2026. The model is designed to reconstruct a 3D scene from a live feed from an ordinary camera, without LiDAR or a depth sensor. It processes frames directly and produces outputs including camera pose estimates, depth data, and point clouds. Available descriptions report a speed of about 20 frames per second on one GPU.¹²

That approach could matter for robotics and spatial applications that need a scene model during capture rather than after lengthy offline processing. The speed figure, however, is supported here by secondary descriptions rather than primary performance documentation. Those accounts do not establish which GPU was used or whether the figure covers model inference alone or the full processing pipeline. Companies should test throughput on their own cameras and scenes before relying on it for real-time requirements.¹

The open-source Colibrì project claims it can run Z.ai’s 744-billion-parameter GLM-5.2 on a machine with about 25 GB of RAM and no GPU.¹² The proof of concept, which became public in July 2026, is a CPU inference engine, not a new model.¹² It keeps part of the mixture-of-experts model in memory and reads the required experts from an NVMe SSD on demand.¹³

This approach reduces RAM requirements but calls for substantial local storage and frequent disk access.¹³ On the developer’s 12-core reference machine, reported generation speed was only about 0.05 to 0.1 tokens per second.⁴ Colibrì thus offers a way to experiment locally with large models. It remains a proof of concept rather than an established production-ready system.²