August 15, 2026 · Saturday

Qwen3.8-27B goes open source

A natively multimodal dense model with 262K context expandable to 1M — outperforming Qwen3.7-Plus on real-world coding and office workflows.

Alibaba Qwen ships open weights for Qwen3.8-27B, a hybrid-attention dense model that fits a single GPU.

Alibaba Qwen kept its promise and shipped open weights for Qwen3.8-27B, a natively multimodal dense model that lands just two weeks after the 2.4-trillion-parameter Qwen3.8-Max. Despite its modest 27B scale, the model outperforms Qwen3.7-Plus across the board, with the biggest gains in real-world coding and office workflows. The architecture is a hybrid-attention dense design: 48 of 64 layers use linear attention, paired with a vision tower and a built-in multi-token-prediction draft head. Native context is 262K tokens and easily extends to 1M.

Day-0 support is already everywhere. vLLM ships single-GPU recipes in BF16 and FP8, SGLang reports 206 tok/s decode on a single RTX 5090, and Ollama lists the model with harness launchers for Claude Code and OpenCode. For a 27B dense model to fit one Blackwell GPU in every precision is the headline in itself — open weights that run where you are, not in someone else's data center.

Claude Code

Auto becomes the default permission mode

Anthropic rolls out Auto mode as the default in Claude Code for Pro, Max, and Team users. If you have already set a default, Claude will ask before changing anything. Shift+Tab still switches modes on the fly, and a defaultMode setting pins one permanently.

Google DeepMind

Gemini 3.7 Flash reaches Pro and Ultra

Gemini 3.7 Flash is now available to all Pro and Ultra users in Gemini chat. The update delivers improved reasoning, framed by GeminiApp as the fastest route to the new Flash tier inside the chat app.

SGLang

206 tok/s on a single RTX 5090

Day-0 SGLang support decodes Qwen3.8-27B at 206.1 tok/s on one RTX 5090 and 38.28 tok/s on DGX Spark — a new bar for what a small model can do on a single card.

vLLM

vLLM lands Day-0 for Qwen3.8-27B

A 27B hybrid-attention model with 262K native context expandable to 1M, ready for single-GPU deployment in vLLM the day it ships.

Photo Story · Z.ai

GLM-5.3, post-trained from the 743B base model, is tuned for code and cyber defense.

GLM-5.3: built to code, ready for cyber defense

Z.ai's GLM-5.3 doubles down on two fronts. It claims top-tier coding and agentic capability achieved through post-training on the 743B base model, plus a major leap in cybersecurity that sets a new standard among open models. Early testers report precise, concise long-task performance, with prefill around 1 ktok/s and output near 60 tok/s.

By next year, using a computer will be optional. Work will radically change.

— Amjad Masad, CEO of Replit

Open Weights · Day-0Model Release

DeepSeek

DeepSeek-V4-Pro ships MIT-licensed

A big jump in agentic capability over the preview, with the same architecture so your config carries over untouched. vLLM has run this path since 0.25.0.

SGLang

dots3-note preview: a 280B agentic model

A 280B-total / 16B-active multimodal model with 512K context, built for real-world long-horizon agent tasks — the model behind RedNote's IMO 2026 gold.

Ollama

Ollama adds the DeepSeek Harness

Run it entirely in your own environment, with web search pre-installed and a trajectory view that shows what is happening in the background.

Anthropic · Policy

Anthropic explains its AI watermarking

Watermarking is implemented to comply with the EU AI Act, and other major model developers signed the same Code of Practice and are following suit.

Anthropic · Safety

The second Risk Report is out

A routine disclosure under the Responsible Scaling Policy, sharing detailed information on system risks and how prepared the company is to address them.

Hugging Face

Small models still dominate local inference

The State of Open Models shows frontier models getting larger while small models dominate real-world use. Qwen leads local inference, followed by Gemma, and AI agents are a major Hub force.

Commentary

Build flexibility into your AI stack

Ethan Mollick warns that extrapolations about AI's corporate impact are too confident, given how unstable today's prices, adoption, and capabilities remain.

Vision models understand the structure of the world projected onto an image far more than LLMs understand the world projected into words — we should tap into the former to do better abstract thinking with the latter.

— François Fleuret

Pika

Four frontier audio models, 20× cheaper

Soundtrack, Music, SFX, and Speech cover the full spectrum of generative sound, priced below every audio model on the market.

Higgsfield

Seedance 2.5 makes 1080p commercials

Full 30-second commercials from a single prompt, with free 1080p generations for new users at zero credit cost.

Perplexity

The Search SDK goes agentic

The SDK behind Perplexity Computer is now available inside any agentic harness, for wide and deep research.

Research

A coding harness nearly solves ARC-AGI-3

Adding a coding harness nearly solves ARC-AGI-3 — coding generalizes LLMs, exactly as predicted.

Tim Dettmers

Testing GLM 5.3: precise and concise

Prefill around 1 ktok/s and output near 60 tok/s. In his harness, full thinking traces already put GLM 5.2 above Fable and Claude Code — GLM 5.3 is another level.

François Chollet

The ARC-3 public set is a demonstration set

Public ARC 3 games are neither eval nor training data; scores on them do not predict scores on the hidden set.

BriefsAround the Industry

Grok