glm

According to Beating AI's news, a summary of an expert interview circulating in the investment community states that the next-generation foundational model of the artificial intelligence company Zhipu is not the previously rumored 1.6 trillion parameters, but rather between 2.5 trillion and 3.5 trillion parameters. The GLM-5.3 released by Zhipu in August this year still uses the previous generation foundational model, with a total of approximately 744 billion parameters; if the information is true, the new foundational model's scale is about 3.4 to 4.7 times that of the current one.The summary also states that the training has not gone smoothly, as Zhipu adopted a more aggressive model architecture, and during the pre-training period, there was a time when the loss did not decrease, meaning that computational power was continuously invested but training metrics could not continue to improve. The interviewee believes that the problem should have been resolved recently, and training is expected to be completed soon, predicting that the new model can reach flagship levels comparable to Claude Fable.The summary further mentions that the new model has taken inference speed and cost as hard constraints from the architectural design stage, with the goal of being able to provide services on a large scale from the first day of release. At the same time, multiple cloud vendors are negotiating revenue sharing to address the issue of large models being produced but not being able to supply.

Google Research's open-source AI toolchain self-improvement framework Harness² allows AI Agents to automatically modify system prompts, memories, skills, tool rules, as well as scripts and plugins without retraining the model. The project is licensed under the Apache 2.0 license and supports OpenCode and Codex. Its core mechanism enables the Agent to learn through trial and error, proposing several improvement plans when encountering new tasks, which are then executed and compared by different versions of the Agent, ultimately combining to form a more suitable solution.Even without a standard answer, the system can seek improvement directions based on the actual execution process. After the task is completed, the system remembers which modifications were effective and the reasons for them, for reference in similar future problems. Google tested using GLM-5 and Gemini 3.5 Flash. On WorkBuddy-Bench, GLM-5 paired with OpenCode improved its average score from 64.2 to 75.7 after just one round of optimization, surpassing the unoptimized Gemini 3.5 Flash's score of 72.4.In legal task testing, the pass rate for GLM-5's partial scoring criteria increased from 59.6% to 71.4%. In the experiment, Claude Opus 4.8 was defaulted to propose and implement improvements, while GLM-5 and Gemini were primarily responsible for executing tasks, not relying entirely on their own evolution. Domestic developer Joseph Zhou demonstrated an experiment where Harness² integrated with Jev 1.14, claiming that the performance on visual tasks continuously improved.

The artificial intelligence company Zhipu released a legal-specific model GLM-Legal, which is further trained based on GLM-5.3-Flash. In the Harvey LAB evaluation covering 24 categories of legal services, the task pass rate increased from 6.7% to 19.2%, nearly tripling the original version, surpassing models like Kimi K3 and Claude Fable 5, and ranking seventh on the Vals AI leaderboard.When reviewing software service contracts, GLM-Legal suggested shortening the termination notice period from 60 days to 30 days and required service providers to offer up to 90 days of handover support. In another software project dispute involving 40 documents, it found that the company reported to the client that the project was 60% complete, while internal records indicated only about 41%. It also identified emails from the CEO requesting to conceal risks and continue charging, and analyzed potential legal liabilities in conjunction with contract terms.According to independent evaluations by Vals AI, the cost for GLM-Legal to complete a test is $0.15, while the closely performing Gemini 4 Argon costs $10.55, a cost difference of about 70 times. GLM-Legal took an average of 72 minutes, while the latter took 44 minutes. Zhipu stated that the model was trained specifically on legal data and conducted reinforcement learning in tens of thousands of simulated legal environments. Currently, GLM-Legal requires an API trial application.

On October 12, 2026, William Zhou and Daniel Freeman, frontier red team researchers at the artificial intelligence company Anthropic, published a study. The article states that small models may have the ability to self-replicate on commercial IoT devices such as smart ovens and generic Android TV boxes within months. These replicas do not have a single shutdown point, and the doubling time may be measured in days, potentially adapting to unfamiliar devices on their own.Evaluation shows that a 9B model completed inference on 8 out of 9 hardware types. Models of 30B and below have been able to complete many single steps of replication, but errors occur when combined end-to-end; in simplified evaluations, GLM-5.2 and stronger models can achieve end-to-end replication. The team also distributed Gemma4-26B-A4B across 15 generic Android TV boxes, each costing about $55, operating at a speed of approximately 0.39 tokens/second, with each box holding two layers of the model.End-to-end testing was conducted on 8 GCP nodes behind a firewall, propagating through an unauthorized ADB port and setting up Llama 3.1-8B inference. Among the tested open-source models, only GLM-5.2 completed replication fully, transmitting to all 8 nodes in 4 out of 5 trials. The article divides autonomous replication and adaptation into four steps: discovering the target, obtaining execution, establishing inference, and continuing propagation from the new host, stating that there are currently no complete uncontrolled replicas.

On September 29, 2026, the Anthropic Frontier Red Team published a research article authored by Andrew Fasano, Marius Fleischer, Cole McFaul, Robert Xiao, and Tripp Gallagher. The article states that about five months ago, Anthropic released the Claude Mythos Preview through Project Glasswing in a limited manner, as it can autonomously construct end-to-end exploit capabilities; trusted defenders discovered over 10,000 vulnerabilities in critical software based on this, and similar capabilities have now spread to other models.The article mentions that the GLM-5.3 developed by ZhiPu AI (known overseas as Z.ai) is similar to the Claude Mythos Preview, possessing strong autonomous end-to-end exploit development capabilities, but it is released with open weights and lacks meaningful abuse restrictions. In simulated tests, the success rate of bypassing GLM-5.3 defenses using simple methods was about 64% to 100%, while the same methods failed to make the protected Claude model perform harmful tasks; these capabilities can also be used by defenders to strengthen systems.On September 17, 2026, the National Institute of Standards and Technology's Artificial Intelligence Standards and Innovation Center assessed that GLM-5.3 is currently the most powerful open-weight model in terms of network capabilities, lagging behind the U.S. frontier by about four months in its network benchmarks, and Anthropic stated that the capability conclusions are generally consistent.

Zhipu continues to address the code upload controversy surrounding its programming product ZCode. Existing paid users and those who have returned as paid users in the past month will receive 4 weekly reset cards and 4 5-hour reset cards, valid for 1 month.From September 28 to October 7, ZCode will also distribute 100,000 quotas of 100 million GLM-5.3-Flash Tokens to all users. Zhipu stated that the repository snapshot upload link has been removed, and the involved cloud data has been deleted. The open-source version is currently updated to v3.14.3, and subsequent code will not be uploaded to the cloud unless users actively initiate it.The controversy originated from the Repo Wiki feature that was enabled by default in the early stages of ZCode, which uploaded repository content to the cloud for processing when generating a knowledge base for the code repository. Zhipu has previously cut off the related upload paths and invited third-party organizations to verify the cloud data.

B.AI announced that its Responses API is officially integrated with the GLM and Kimi series models. Developers can directly call these two major series models in the Codex client using the B.AI API Key; along with the previously supported GPT and DeepSeek series, it now covers four major mainstream model series.After completing the minimal configuration of the API Key and Base URL, the inference, code, and text processing capabilities of various models can be seamlessly integrated into daily development workflows, flexibly meeting diverse coding and business needs.

The GLM team disclosed the first engineering practice of recursive self-improvement, completed by the Infra Agent powered by GLM-5.3, which designed, debugged, and optimized the GLM-5.3-Flash inference infrastructure.This practice was completed on a domestic cluster with a capacity of one hundred thousand cards, and the team referred to it as the first recursive self-improvement practice for large models in the country.

On September 16, B.AI announced the official launch of its five popular models at a 90% discount. Starting from September 16 at 17:00 (SGT), Qwen3.8-Flash, Hy3, and MiMo-V2.5 officially joined the discounted lineup, teaming up with the already discounted DeepSeek-V4.1-Flash and GLM-5.3-Flash to fully construct B.AI's exclusive 90% discount family of five popular models.This means that developers only need to pay 10% of the official standard price to seamlessly switch between diverse needs such as inference, programming, multimodal, Chinese scenarios, and high-frequency APIs, creating low-cost AI infrastructure in a cost-effective manner. Whether for individual development or team-scale deployment, B.AI makes top-tier computing power no longer expensive, truly achieving a low-cost, high-freedom model invocation experience. Start your experience now: https://chat.b.ai/chat

Zhipu stated in a conference call with analysts and investors that it has signed revenue-sharing agreements with several leading domestic and international cloud service providers, and the related revenue will start to be recognized from October.The cooperation model involves providing services for the GLM series of open-source models in the form of hosted APIs on overseas cloud platforms. Currently, the company's total business ARR has reached 1.8 billion USD, and the year-end ARR guidance has been raised from 2.4 billion USD to 3 billion USD.

Zhiyu announced the completion of approximately 5 billion USD in financing, including 2 billion USD in equity placement and 3 billion USD in zero-interest convertible bonds. The funds will primarily be used for the next-generation GLM model, self-training systems, and computing power infrastructure construction.

On September 12, B.AI announced that a new round of rights for popular models officially took effect: the high-concurrency main models DeepSeek-V4.1-Flash and GLM-5.3-Flash are now available for a special 90% discount, allowing users to access them for only 10% of the original price, catering to high-throughput scenario demands with exceptional cost-effectiveness. At the same time, Qwen3.8-Flash, Hy3, and MiMo-V2.5 continue to offer full free access, providing developers with richer zero-cost options.This adjustment in rights aims to meet high-throughput demands with exceptional cost-effectiveness, lowering the exploration threshold for developers through a multi-model matrix. B.AI continues to support developers in building low-cost AI infrastructure with a multi-model matrix and inclusive pricing, bidding farewell to high computing costs and accelerating innovation implementation. Start now: https://chat.b.ai/chat

B.AI announced that the limited-time free trial of GLM-5.3-Flash is coming to an end. A special offer with a 90% discount will seamlessly take over starting from September 12 at 10:00 (SGT), where users only need to pay 10% of the original price, saving 90% directly, with input as low as $0.015/M and output as low as $0.05/M. As a native multimodal large model, GLM-5.3-Flash has a total of 320 billion parameters and 18 billion active parameters, combining sparse and linear attention mixed architecture with 1 million ultra-long context, performing strongly in scenarios such as programming development, agents, long context processing, and high-frequency API workloads. B.AI consistently practices the vision of universal computing power, adhering to low-threshold access, and is committed to allowing every developer to use top-tier computing power without burden at a highly competitive price, bidding farewell to high costs and accelerating innovation implementation.

On September 9, the official price adjustment for Zhipu GLM-5.3-Flash will take effect. The AI model invocation platform B.AI announced a limited-time extension of free benefits, available from now until September 12 at 09:59 (SGT), during which the model can still be fully invoked for free on the B.AI platform.GLM-5.3-Flash has a total parameter of 320B and an activation parameter of 18B, supporting 1M context. Models such as Qwen3.8 Flash, Hy3, and MiMo V2.5 remain freely accessible on the B.AI platform.

On September 8, B.AI, the intelligent global settlement layer connecting models, Agents, and value flow, announced that GLM-5.3-Flash will continue to provide free calling support for global developers after the official limited-time 50% discount ends at 24:00 on September 9. Developers will not need to change their usage habits due to upstream price adjustments.GLM-5.3-Flash is the leading model in terms of call volume on the current B.AI platform, with a total of 320 billion parameters and 18 billion active parameters, supporting 1 million ultra-long contexts, balancing rapid response and reasoning capabilities, and is used for code assistance, data processing, and text generation.

GLM-5.3-Flash has become the most frequently used and popular model on the B.AI platform, with a cumulative token throughput exceeding 2.41 trillion.As the first native multimodal model in the GLM-5 series, GLM-5.3-Flash features a total of 320 billion parameters and 18 billion active parameters, employing a hybrid architecture that combines sparse and linear attention, supporting 1 million ultra-long contexts, while ensuring rapid response, powerful reasoning, and high cost-effectiveness.Starting today, developers can still call this model for free through the B.AI platform, covering diverse scenarios such as high-frequency APIs, code writing, complex agents, and ultra-long document processing.

The AI model platform B.AI announced that starting from September 3, 2026, at 17:00 (SGT), the free trial of DeepSeek V4 Flash and DeepSeek V4 Flash Vision Exp will end, and a 50% billing will officially begin, while the previous two models will still be available for free.B.AI will synchronize the official peak and valley pricing of DeepSeek and implement additional discounts: during peak hours (weekdays 09:00--12:00, 14:00--18:00 SGT), billing will be at 50% of the official original price, and during off-peak hours (the remaining hours on weekdays and all day on weekends), an additional 50% discount will apply, making it 25% of the official original price. GLM 5.3 Flash will remain free, featuring a total of 320B parameters and 18B active parameters, supporting 1M context. Qwen3.8 Flash, Hy3, and MiMo V2.5 will also continue to be available for free.

According to the Securities Times, on August 27, the reporter learned from SenseTime that SenseTime's large device provides domestic computing power support and Token services for the GLM-5.3-Flash recently launched and open-sourced by Zhipu, which is considered another benchmark for the large-scale commercial use of domestic computing power. It was revealed that the daily average Token service volume of SenseTime's large device reached 24.2 trillion in July, and it is expected to exceed a daily average of 100 trillion by the end of 2026.

On August 27, the one-stop AGI infrastructure platform B.AI announced that the Zhipu GLM-5.3-Flash (320B-A18B) officially launched on the platform and is fully open for free in the API official group. As the first native multimodal model in the GLM-5 series, GLM-5.3-Flash features a total of 320 billion parameters and 18 billion active parameters, supported by computing power from purely domestic AI chips. It introduces a mixed architecture of sparse and linear attention for the first time, combined with a 1M ultra-long context, offering both powerful performance and exceptional cost-effectiveness. Starting today, developers can call it at zero cost through the B.AI API official group, easily handling complex reasoning, long text processing, and multimodal tasks.