GoKawiil
Tech News
Artificial Analysis has begun benchmarking small AI models that run directly on mobile phones, focusing on models that fit within 8 GB of memory after quantization, including KV cache at 8K context. The effort combines intelligence benchmarks with real on-device inference data gathered in partnership with Liquid AI, whose measurement methodology Artificial Analysis says it has independently validated.
An unidentified free model called Ox Alpha appeared on OpenRouter and quickly drew massive usage from developers who couldn't determine its origin, sparking days of speculation about which lab built it. Z.ai revealed the model was actually its own GLM-5.3-Flash, running entirely on Chinese chips, and priced at 15 cents per million input tokens and 50 cents per million output tokens, with a launch discount cutting that in half through September 9.
OpenAI has revealed details of Jalapeño, a custom inference chip developed with Broadcom over roughly 16 months, at the Hot Chips conference. SemiAnalysis tested the chip using its InferenceX benchmark suite and found it outperforms Nvidia, AMD, and Google chips across multiple open-source models, despite being OpenAI's first hardware effort.
At Hot Chips 2026, IBM unveiled an early look at a future mainframe processor featuring 11 IBM Z cores on a 2-nanometer process, clocked above 5.7 GHz, capable of natively executing both z/Architecture and AArch64 instructions on the same cores without emulation. The chip will include specialized accelerators for AI fraud detection, I/O, compression, cryptography and sorting, plus a large cache hierarchy topping out near 3.5 GB of virtual L4 cache. IBM hasn't named the chip or its host system but expects it around 2028, likely powering a future z18-class mainframe.
Researchers documented that vLLM's XML-based tool parser for Qwen3 Coder (CVE-2025-9141) fed model-generated arguments directly into Python's eval() function, meaning a language model's own output tokens could execute arbitrary code on the machine running it. A Gemini-based review flagged the change as a critical risk before it was merged, but the lead vLLM maintainer force-merged it anyway.
OpenAI is temporarily lowering the price of its GPT-5.6-Sol model, with the discount running until at least November 21. Separately, the company is phasing out its fine-tuning platform, closing it to new users while letting existing customers keep submitting training jobs for a limited time. Models already fine-tuned will still work for inference until their underlying base models are retired.
Anthropic published a terminology guide clarifying that an 'agent system' is not the same as a 'model.' It breaks the stack into three parts: the model (like Sonnet or Opus), the inference service that runs it (like AWS Bedrock or Anthropic's API), and the harness, the interface layer such as Claude Desktop or Claude CLI that manages tools and context.
A developer created Varkos, an AI-driven companion mod for Skyrim that can fight, loot, fetch items, and carry out multi-step commands in real time using local inference instead of cloud-based LLM calls. The project aims to fix the latency and weak world-interaction problems common in existing AI NPC dialogue systems, while giving Varkos a persistent personality that evolves with shared gameplay experiences.