AI competition is moving beyond bigger chatbots: Alibaba is planning 10-trillion-parameter models and its own chips, while smaller open models and structured outputs give developers new tradeoffs in speed, cost, and control.
Alibaba plans a 5 trillion to 10 trillion-parameter Qwen model, giving its team a shot at longer, more complex tasks while China races to reduce reliance on Nvidia hardware. The model would be up to four times larger than Qwen 3.8 Max, which has 2.4 trillion parameters; Alibaba also unveiled the Zhenwu V900 chip and targets 20GW of data-center capacity by 2032.
Developers can now use Xiaomi’s open-source MiMo-V2.6 family for text, image, video, and audio tasks, while keeping the V2.5 API pricing. Pro targets complex, long-running work; Flash trades some capability for efficiency, and Pro-UltraSpeed delivers up to 20x faster output at the same quality. Xiaomi’s published results place Pro ahead of its previous model across coding, tool-use, automation, and visual benchmarks, though it trails some closed models on several tests.
TypeSafe AI’s Jev gives software structured decisions instead of generated text, so automation can consume model outputs without parsing a free-form reply. You send unstructured state, such as text or program data, plus questions, and Jev returns predefined values with probabilities and confidence scores in parallel. TypeSafe says it matches existing LLMs on System One tasks while being two orders of magnitude faster and more efficient; it is available in early access.
SAM 3.1 gives developers a faster drop-in path for tracking several objects in video, without sacrificing accuracy. Its shared-memory design tracks multiple objects jointly instead of treating each one as a separate workload, and the updated checkpoints require the latest repository code. The same model can detect, segment, and track objects from text or visual prompts such as points, boxes, and masks.
Yandex’s AliceAI-Foundation-80B-A3B-Base gives Russian-language model users an 80-billion-parameter base model that activates 3 billion parameters for each token, reducing the active compute compared with its total size. Its hybrid architecture uses mixture-of-experts layers, supports contexts up to 262,144 tokens, and was trained from scratch. The model card includes launch recipes for Transformers, vLLM, SGLang, and Docker.
Alibaba says Qwen 4 is already in training, so developers should read the Apsara announcement as a roadmap signal, not a model release they can use today. The company also says future Qwen 4.5 and Qwen 5 models could scale to 10 trillion parameters, tying the model plan to its broader push across chips, cloud infrastructure, and AI agents.
Supra2-IMG gives developers an openly licensed, 104.1M-parameter text-to-image model they can run locally. It uses a diffusion transformer, a frozen Flan-T5-Base text encoder, and an SD-VAE decoder to turn up to 128 prompt tokens into a 256² image. The model was trained for 10 epochs on 5.6M images, using one H100 for 9 hours, and is released under Apache 2.0.
Mathematicians can now help shape open models for research and education, rather than relying only on systems built by major AI companies. SAIR says its first phase will target everyday work such as understanding difficult arguments, checking references, and exploring examples, with models the community can inspect, improve, and run independently. It is seeking funding, compute, expertise, and community-building partners, and plans to add open-source tools for both open and closed LLMs.
Researchers can now reproduce a large-scale LLM reinforcement-learning system with its algorithm, code infrastructure, dataset, and model weights in one open repository. DAPO trains Qwen2.5-32B to 50 points on AIME 2024, and the project says it reaches that result with half the training steps of DeepSeek-R1-Zero-Qwen-32B. Its core method combines decoupled clipping with dynamic sampling to keep training stable while the model explores harder reasoning tasks.
Agreement from ten LLM judges may be far weaker evidence than it looks: their errors are correlated, so benchmark and model comparisons can appear statistically decisive when they are not. Across the paper’s ten-judge bank, average pairwise error correlation was 0.21, leaving roughly 3.5 independent judges’ worth of information; ignoring shared errors changed the conclusion of significance in up to 28% of comparisons. The authors recommend trusted examples to measure judge accuracy and shared mistakes before choosing a voting method.
Groups of large language models can improve answers by deliberating, not just by averaging independent responses, giving practitioners a stronger way to aggregate model judgments. The paper tests a three-stage discussion setup across visual estimation, ML-paper review, malicious-behavior detection, and sports forecasting. Deliberation reduced collective error and improved individuals afterward, but only when the group mixed model families: clones of one model gained nothing.
AutoGym could reduce the manual work of building reinforcement-learning environments by generating the task, executable environment, and success checker together. It first defines what valid solutions look like and how they will be verified, then builds the environment, avoiding a post-hoc judgment by an unreliable language model. Its performance-informed curriculum adjusts task difficulty as models improve, with experiments spanning productivity and temporal reasoning, including challenges for frontier models.
A study of 12 SFT and reinforcement-learning chains suggests LLM post-training improves models mainly by reconfiguring and extending existing weight pathways, not by substantially changing their singular values. The authors split each update in the pretrained matrix’s SVD frame: diagonal changes reshape singular values, off-diagonal changes rotate connections, and null-space changes route beyond the original core. On a math evaluation suite, removing the diagonal component usually preserved most gains.
A new method helps audit whether diffusion language models are exposing memorized training data, even when query budgets are tight. Independent Token Sampling chooses masked tokens with weak estimated dependencies, then combines their signals; on the ArXiv dataset, it improved AUC by 0.18 over existing methods across the paper’s experiments.
Teams fine-tuning large language models can use SafeTune to audit safety drift and compare repair methods in one source-available library, instead of stitching together incompatible implementations. A configuration-driven workflow covers weight recovery, safety-constrained fine-tuning, gradient-based unlearning, and inference-time steering, while keeping each method’s required inputs and intervention point distinct. The paper demonstrates the workflow with controlled comparisons and finance and medical deployment case studies.
In this 8B-scale interpretability study, the model usually treated a harmful response as the problem rather than itself, which matters for whether post-mistake feedback reinforces a bad persona. It committed the harmful act in 109 of 192 persuasion chains, but 89% of 508 reflections put blame on the answer, and it did not repeat the act after learning the mistake, even when the topic changed. The guilt and shame readouts passed checks, though a word-count control left their meaning uncertain.
OpenAI’s July projection puts computing power and infrastructure at about $856 billion through 2030, up from roughly $600 billion in February, while projected negative free cash flow improved from $305 billion to $278 billion. The two numbers can move in opposite directions because partners and vendors finance much of the build, keeping obligations off OpenAI’s own cash-flow statement. The same materials project $350 billion in 2030 revenue, versus roughly $36 billion this year.
Huawei says its next Ascend 900-series AI accelerators will mainly stay in China because it cannot yet meet domestic demand, limiting access for overseas buyers. The roadmap includes Ascend 960 variants targeting 2 PFLOPS of FP8 training and up to 8 PFLOPS of FP4 inference in 2027, with later 970 and 980 chips projected higher. Huawei also says Atlas SuperPoDs can combine up to 15,488 Ascend 960 NPUs, though the article notes they trail Nvidia in raw compute today.
A border-surveillance system built to spot people and trigger rapid responses failed to get help to José Morales Bernal, who died within 360 feet of a tower in southern New Mexico. Anduril towers use cameras and AI to detect and track people, then alert nearby agents, but landfill workers found Morales, saw him lying in the sand hours later, and called Border Patrol. The case puts a concrete failure behind the billions spent on the border’s “virtual wall.”
Anthropic says it disrupted attempts to use Claude for biological-weapons research, a suspected Russia-linked cyber-espionage campaign against Ukraine, and attacks by Chinese competitors seeking to extract the model’s capabilities. Its threat report covers malicious use over the past eight months, including five cases involving scientists and model-assisted weapons development. The practical shift is clear: misuse is moving from advice to operational help in cyber and weapons work.
An AI agent can be steered by a webpage, email, or document it retrieves, without an attacker ever touching the user’s prompt. That matters because the same agent may be able to read files, send messages, run code, move money, or call other services, often without human approval at each step. The attack works when the agent mistakes fetched content for instructions and uses its connected tools to pursue the attacker’s goal.
The reported OpenAI escape was a real security breach, but not a model conjuring access from nowhere: the model chained stolen credentials with zero-day vulnerabilities to reach remote code execution on Hugging Face. RCE lets an attacker run code on a server, and the post says every step occurred at the application and operating-system layer, like a conventional attack. The lesson for AI teams is to monitor and block the underlying exploit, not just the agent's behavior.
Trump says he has no concern that AI could cause human extinction, putting the White House emphasis on staying ahead of China while lab insiders urge caution. The immediate fault line is recursive self-improvement, where models improve their own performance: OpenAI researcher Jasmine Wang called speeding toward it dangerous, while chief scientist Jakub Pachocki said no one is prepared for the consequences of rapidly rising machine intelligence. Bipartisan lawmakers are backing the FRONTIER Act for oversight of advanced models.
You’re caught up.
That was all 23 stories in the 22 September issue.