Beacon

Latest updates from supported sources.

Tencent Hy News · 2027-08-05

From LR to ELR: A Better Heuristic for Pretraining Dynamics

Understanding pretraining dynamics is crucial for designing effective training hyperparameters for large language models (LLMs), particularly the learning rate (LR). However, LR does not always reflect how much the function represented by the model actually changes, and the optimal LR often shifts with model and data scale. In this post, we identify that the effective learning rate (ELR), which controls the directional changes of model weights, is a more intrinsic heuristic than LR as a tunable hyperparameter. Specifically, ELR delivers more accurate loss prediction under the multi-power law (MPL) model and transfers more reliably across model scales. More broadly, the ELR perspective guides the design of better schedules, which outperform conventional LR-schedule baselines.

Open original

Kiro Changelog · 2026-09-14

Models: GPT-5.6 Sol, Terra, and Luna upgraded to 1M context window

GPT-5.6 Sol, Terra, and Luna now support a 1M token context window, up from 272K, in the Kiro IDE, CLI, and Web. Entire codebases, long-form documents, and extended multi-turn agent histories fit in a single request, so repository-scale refactors, end-to-end debugging, and long-horizon agentic tasks run without chunking or losing earlier context. Credit multipliers for the GPT-5.6 family are updated with this launch and follow the two-tier context model OpenAI and Amazon Bedrock use: requests up to 272K tokens bill at the short context rate (Sol 4.4x, Terra 2.2x, Luna 1.1x), and requests above 272K tokens bill at double that rate. 1M context is rolling out gradually with experimental support to Kiro Pro, Pro+, Pro Max, and Power customers. Restart your IDE or CLI, or refresh Kiro Web, to see the updated models. Learn more ->

Open original

Kiro Changelog · 2026-09-14

IDE: Agent Artifacts and Native ARM64 Builds

IDE 1.1 lets agents produce durable artifacts you can review and return to, adds native ARM64 builds for Windows and Linux, and updates the editor foundation to Code OSS 1.131. Compacted conversations retain the right context, MCP failures are clearer, Hooks are more reliable, and enterprise profiles route requests to the selected region.

Open original

Artificial Analysis Changelog · 2026-09-13

K2 Horizon 7B

New language model evaluation results available — Intelligence Index: 21

Open original

Artificial Analysis Changelog · 2026-09-13

K2 Horizon 3.7B

New language model evaluation results available — Intelligence Index: 16

Open original

Artificial Analysis Changelog · 2026-09-13

K2 Horizon 0.9B

New language model evaluation results available — Intelligence Index: 3

Open original

OpenRouter Blog · 2026-09-11

Zero Data Retention (ZDR): What It Means for AI APIs

ZDR means an AI provider processes your prompt, returns a response, and doesn't store either one afterward. It's a retention guarantee, not a universal privacy policy. This page explains the boundary, compares ZDR with related controls, and shows how to enforce it at the account, guardrail, or request level.

Open original

OpenRouter Blog · 2026-09-11

OpenRouter Text-to-Speech: API Tutorial in 5 Minutes

Our OpenAI-compatible speech endpoint puts TTS models from Mistral, xAI, Microsoft, and more behind one request shape. Here's the path from API key to a playable MP3 in cURL, Python, JavaScript, and the OpenAI SDK, plus the response checks that keep JSON errors out of your audio files.

Open original

Artificial Analysis Changelog · 2026-09-10

Ling-3.0-flash-VL

New language model evaluation results available — Intelligence Index: 25

Open original