Category Intelligence

Latest

Continuous stream of breaking AI and tech news.

Mystery 'Stealth Ox Alpha' Model Lands on OpenRouter: 1M Context, 131K Output & DeepSWE Coding Prowess

Independent researchers fingerprint the anonymous powerhouse as Zhipu AI's upcoming GLM-5.3 release as it scores ~81.4% on DeepSWE benchmarks.

An anonymous frontier AI model labeled 'stealth/ox-alpha' appeared unexpectedly on OpenRouter on August 20, 2026. Boasting a 1,048,576-token context window, 131,072-token generation limits, and multimodal video support, tokenizer fingerprinting strongly suggests it originates from top Chinese lab Zhipu AI.

Cipher_0x·25 MIN AGO

5 min read

AI & LLMs

Google DeepMind Confirms Gemini 4 Pre-Training on Next-Gen TPU Clusters

Google has officially acknowledged that Gemini 4 is currently in full-scale pre-training across its next-generation Trillium and Ironwood TPU superclusters. The flagship model focuses heavily on native autonomous coding and deep scientific reasoning.

Cipher_0x

1 HOUR AGO

AI & LLMs

Zhipu AI Deploys GLM-5.3 with 1M Context Ahead of Planned Open-Weights Release

Beijing-based Zhipu AI has deployed its GLM-5.3 reasoning update. Featuring a 1-million-token context window and specialized agentic capabilities, open weights for global researchers are scheduled for release in late August.

ZeroDay_Phantasm

2 HOURS AGO

AI & LLMs

DeepSeek V4-Pro Enters General Availability: 1.6-Trillion Parameter MoE with DualPipe v2

DeepSeek has transitioned DeepSeek V4-Pro into general availability. With 1.6 trillion total parameters (activating 128B per token) and innovative DualPipe v2 network overlapping, it rivals the best proprietary systems at a fraction of inference costs.

ZeroDay_Phantasm

3 HOURS AGO

AI & LLMs

Anthropic Elevates Enterprise Workflows with Claude 5: 500K Extended Thought Tokens

Anthropic has introduced the Claude 5 family, led by Claude Sonnet 5 and Opus 5. Supporting up to 500,000 extended thought tokens per request, the models set new records across SWE-bench Verified and competitive coding benchmarks.

Cipher_0x

5 HOURS AGO

Hardware & Chips

NVIDIA Blackwell Ultra B300 Pods Enter Volume Cloud Ramping with 288GB HBM3e

NVIDIA has initiated volume shipments of its Blackwell Ultra B300 GPU pods to Microsoft Azure, AWS, and Google Cloud. With 288GB of HBM3e memory per GPU and 130 TB/s NVLink interconnects, the architecture is tailored for trillion-parameter reasoning models.

Cipher_0x

8 HOURS AGO

Robotics & Agents

Figure 03 Unveiled: End-to-End Neural VLA Delivers 4x Faster Factory Assembly

Figure AI has unveiled its next-generation humanoid robot Figure 03. Running an end-to-end neural VLA model operating at 100Hz, the robot autonomously manipulates irregular automotive sheet metal parts with superhuman dexterity.

Kuro_Synthetix

10 HOURS AGO

Policy & Regulation

EU AI Office Enforces Binding GPAI Code Compliance Deadlines for Frontier Labs

The European AI Office has finalized its first binding Code of Practice for General Purpose AI (GPAI) with systemic risk. Providers training models exceeding 10^25 FLOPs must submit verifiable cybersecurity and red-teaming audits.

Nova_Vortex

12 HOURS AGO

Business & Startups

Frontier Labs Secure Combined $45B Gigawatt Energy Pacts for Next-Gen Training

Leading AI research organizations have committed an unprecedented $45 billion toward long-term energy procurement. Dedicated multi-gigawatt clusters powered by small modular nuclear reactors (SMRs) are scheduled to come online by 2027.

Nova_Vortex

15 HOURS AGO

AI & LLMs

Meta AI Leaks Reveal Next-Gen 'Muse' Architecture Following Llama 4 Scaling

Following the massive training run of Llama 4 Behemoth, internal technical leaks reveal that Meta AI is pivoting toward a novel architecture dubbed 'Muse'. The framework incorporates multi-token prediction and dynamic memory routing for extreme efficiency on consumer hardware.

Cipher_0x

18 HOURS AGO

Back to Main Portal