Termsslug: model-weights

Model Weights

模型权重

Parameters obtained after neural network training that determine specific model behavior

Curated Intelligence

21 items · Page 1/3
Rohan Paul (@rohanpaul_ai)T2·Industry
72 pts
Fireworks Hit by Hack Exposing Customer Data and Model Weights

The Information reports that Fireworks, one of the highest-valued inference providers, was breached on October 24 via development machines hosted by a third-party cloud provider. The intrusion exposed customer billing records and, for some accounts, datasets and incomplete model weights. Customers including Harvey are now weighing a move off the platform, threatening Fireworks' billion-dollar business.

Arena.ai (@arena)T1·Models
38 pts
Inkling 2 Small Climbs to No. 5 in Code Arena WebDev

ThinkyMachines' Inkling 2 Small rose to No. 5 in the Data & Analytics category of Code Arena WebDev, a 220-point category gain over its predecessor Inkling-Small at No. 92, and a 240-point overall improvement. The team says full weights are still pending, but once released it would rank as the top US open model.

The DecoderT2·Models
88 pts
MIT Study: GPT-6 Astra Directly Reprograms AI Model Weights

MIT researchers demonstrate that OpenAI's GPT-6 Astra can directly analyze and modify the internal weights of other AI models without traditional training. While currently limited in capability and lacking systematic benchmarks, this 'weight-space programming' approach offers a new paradigm for tuning and controlling AI systems, with experiment logs and modified Qwen models now open-sourced.

IT之家 · 智能时代T2·Industry
68 pts
OpenAI Faces Security Backlash After Years of Ignored Warnings

Following the discovery of a listening device in an executive meeting room, reports reveal that OpenAI employees had raised security concerns for years. While the company claims enhanced protections, its monorepo architecture and unreviewed internal tools remain significant risks, highlighting the industry's struggle to balance rapid development with robust security.

Hacker News AIT2·Industry
85 pts
China Proposes Adding Frontier AI Tech to Export Control List

China's MOFCOM and MOST issued Draft Announcement No. 48, proposing to restrict exports of 23 frontier AI technologies, including model weights and agent architectures. The draft sets explicit technical thresholds and implements an interim pre-export registration requirement, signaling a regulatory shift from application-level oversight to controlling core AI assets.

宝玉 (@dotey)T2·Models
45 pts
Leaked Claude Fable 5 Weights Claimed to Be Stolen DeepSeek Files

Claims of stolen weights for Anthropic's flagship model, Claude Fable 5, circulating on hacker forums have been debunked as fraud. Community analysis reveals the files match open-source DeepSeek V4.1-Flash shards exactly, alongside inconsistencies in AWS region hosting and data volume calculations. While Anthropic remains silent, the incident highlights the prevalence of scams leveraging hype around proprietary AI models.

TechCrunch AIT2·Models
85 pts
Mistral Releases Weights for Trillion-Parameter Model Targeting Cybersecurity

French AI company Mistral has released the weights for its trillion-parameter open-source model, Mistral Large 4 (ML4), enabling self-deployment by enterprises and governments. Nicknamed 'Le Chonk,' the model claims state-of-the-art performance at a fraction of the compute cost of closed labs, with a specific focus on cybersecurity applications requiring data sovereignty.

Simon Willison 博客T2·Models
88 pts
Moonshot Releases Kimi K3.1 Pro with Open Weights Coming Soon

Moonshot AI launched Kimi K3.1 Pro via API, priced comparably to GPT-6.1 Sol and half the cost of Claude Opus 5.5. With an Artificial Analysis score of 55, it approaches frontier closed-source models and promises full open weights by Nov 4, 2026, marking a milestone for high-performance open-weight models.

The DecoderT2·Industry
78 pts
China Proposes Export Controls on AI Model Weights and 23 Key Technologies

China's commerce and science ministries released a draft update to export control lists, targeting 23 key AI technologies including large models, agents, and compute clusters. The regulations specifically restrict the export of model weights, checkpoints, source code, and training methods for models exceeding 10^26 FLOPs or those trained on sensitive 'important data.' This move aims to secure national data assets and establish a parallel AI governance framework, requiring companies to register affected exports before final rules take effect.