Termsslug: post-training

Post-training

后训练

Training stage after pretraining that aligns model behavior and boosts specific capabilities with extra compute and data.

Curated Intelligence

51 items · Page 1/6
Hacker News AIT2·Industry
78 pts
Palo Alto Networks to Acquire Applied Compute for $4.5B

Palo Alto Networks has agreed to acquire Applied Compute, a post-training and inference platform for open-weight AI models, for $4.5 billion, bringing its research team and platform into the security vendor. The deal targets hybrid enterprise AI, letting customers securely train, adapt and deploy open-weight models across public and sovereign clouds. It marks a major security-vendor push into AI model infrastructure.

IT之家 · 智能时代T2·Industry
38 pts
Moonshot Executive Says AI Iteration Speed Will Keep Accelerating

Xu Tiantian, Moonshot AI's head of finance industry and ecosystem, told the Caixin Summit that AI intelligence is far from its ceiling, with three growth curves—pretraining compute, post-training and inference, and agent design—compounding to accelerate iteration. She added that agents can independently handle ever-longer tasks, pushing AI from odd jobs into deeper workflows.

IT之家 · 智能时代T2·Models
78 pts
Baidu Releases ERNIE 5.5 Flash Preview with 280B Parameters and Open-Weight Pledge

Baidu has released ERNIE 5.5 Flash Preview, the first post-trained variant of its Base model, featuring 280B total parameters, 14B activated parameters and a 1M-token context window, targeting coding, 3D modeling and content creation. The API is live on the Qianfan platform, with weights promised to be open-sourced by November 25; Baidu describes it as an intermediate post-training checkpoint still under rapid iteration.

AWS Machine Learning BlogT1·Models
62 pts
Moonshot AI Brings Kimi K3.1 Pro Highspeed to Amazon Bedrock

Moonshot AI has released the Kimi K3.1 series, with Kimi K3.1 Pro Highspeed now available on Amazon Bedrock. Built on the same 3T-class open-weight architecture and 1-million-token context window as Kimi K3, it was post-trained with substantially more compute to target long-horizon agentic work, reduce factual errors, and improve professional tasks such as financial modeling, office documents, and coding for game engines, 3D scenes, and CAD scripts. The Highspeed variant delivers 2-3x the output speed of K3.1 Pro and reuses the same endpoints and APIs, so developers can switch by changing the model ID.

Hacker News AIT2·Models
62 pts
Nous Research Post-Trains Qwen3.8-27B into Arete-27B

Nous Research post-trained Alibaba's Qwen3.8-27B to reinforce Western values, releasing Arete-27B with open weights, the Propaganda_Eval benchmark and the arete-data-pairs training set. The team argues that frontier labs share a homogeneous worldview that injects political bias into agents, and that open models let users reshape alignment to their own preferences.

InfoQ 中文T2·Industry
38 pts
Tencent Cloud Restructures Hunyuan R&D Around Real-World Business Evaluation

At its Digital Ecosystem Summit, Tencent Cloud outlined a restructuring of Hunyuan's R&D pipeline: a reorg began in December 2025, post-training infrastructure was upgraded in January 2026 for long-horizon agent reasoning, and major releases now land every two to three months. On evaluation, it supplements public benchmarks with blind tests by 163 internal domain experts and an internal suite spanning nine business lines, addressing the coding-heavy bias of existing agent benchmarks.

量子位官网T2·Models
72 pts
Tencent Releases Hunyuan Hy4, Scales WorkBuddy Tenfold as Demand Surges

Tencent released Hunyuan Hy4, its latest flagship LLM, adding vision understanding for the first time and substantially scaling pretraining and post-training. First-day demand overwhelmed capacity, prompting a tenfold expansion of the WorkBuddy inference cluster and higher free quotas. Hy4 ranks second among Chinese models on Artificial Analysis' intelligence index and first among open models on Agent Arena, at roughly $0.78 per task.

The DecoderT2·Models
78 pts
Baidu Open-Sources 280B-Parameter ERNIE 5.5 Flash Base Foundation Model

Baidu released ERNIE-5.5-Flash-Base on Hugging Face, a mixture-of-experts foundation model with 280B total parameters and 14B active per token, pre-trained on 33.5 trillion tokens with a 128K training context that can extend to one million tokens during fine-tuning and inference. It ships deliberately without fine-tuning or post-training, positioning the raw weights for developer customization, with a post-trained version to follow.

IT之家 · 智能时代T2·Models
58 pts
Tencent Hunyuan Hy4 Demand Surges, Inference Cluster Scaled Tenfold

Tencent Hunyuan's Hy4 hit quota limits on its first day of launch inside WorkBuddy, prompting a tenfold emergency expansion of its inference cluster and higher free-tier quotas, while Hy3's free access was extended to November 15. Compared with the preview, Hy4 substantially scales pre- and post-training, adds vision for the first time, and is billed as the company's first entry among top open-source models, with weights and a system card promised soon.

机器之心 (微信公众号)T2·Agents
58 pts
Qwen Outlines Personal Agent Architecture Built on Model, Memory and Ecosystem

Jiang Guanjun, AI algorithm lead of Qwen's business unit, framed the Personal Agent stack around three pillars: model, ecosystem and memory, with a "demand-driven memory system" that infers what to remember from how users actually use the agent. The team built a demand map spanning 30-plus industries and decomposed 8,000-plus atomic model capabilities, using post-training and reinforcement learning to stabilize task execution.