Qwen fits a 27B multimodal model on a single GPU

Open models

Qwen's 27B vision-language model runs on a single GPU¹

Alibaba released Qwen3.8-27B under the Apache 2.0 license, a 27-billion-parameter dense model that reads text, images, and video. Independent trackers put the release on 14 August.² Three of every four transformer blocks use Gated DeltaNet linear attention, with full Gated Attention only in the fourth, so the memory cost stays flat where a standard stack grows with sequence length. That is what lets the model hold 262,144 tokens natively and stretch to about a million.¹ At 8-bit precision its 27 billion parameters take about a byte each, so the weights come to roughly 27 GB¹—an estimate, not a measured figure. That is still small enough to keep the model on the consumer hardware independent trackers say it runs on,² not a cluster.

Qwen's own evaluation table reports high agentic and coding scores, 61.7 on SWE-bench Pro and 84.3 on OSWorld-Verified.¹ Those are the lab's figures. No independent evaluator had posted verified results by 16 August. The change is in what it takes to run a model this capable and this multimodal: one GPU, not a cluster.

AI provenance

Anthropic will watermark all of Claude's text output³

Anthropic said it will embed an imperceptible statistical watermark in every piece of text Claude generates, using a version of Google DeepMind's SynthID-Text approach, and attach signed C2PA content credentials to the files it produces. The text mark biases word selection from a secret key and the preceding words, so no single passage looks altered while the pattern is recoverable across enough tokens. The driver is regulatory. The EU Code of Practice on Transparency of AI-Generated Content, signed in July 2026 under the EU AI Act, forces it, with a phased rollout for models launched before 2 August 2026.⁴ Anthropic is explicit about the limit: the watermark shows only that Claude was “likely involved,” and it “cannot distinguish ‘Claude wrote this’ from ‘Claude heavily edited this.’” Provenance now rides inside the text across a vendor's entire surface, with no opt-out, where before it depended on a label a user could delete.

Coding models

GLM-5.3 gets its coding leap from post-training alone⁵

Z.ai, formerly Zhipu AI, released GLM-5.3 and says the only change from GLM-5.2 was more post-training. “Scaling post-training is all we did for GLM-5.3,” the lab wrote.⁶ The base is the same 743-billion-parameter model, with reinforcement learning on top and a roughly 50% coding gain over the prior version by Z.ai's count. The positioning is unusual for a frontier model, “Built to Code. Ready for Cyber Defense.” Z.ai reports strong security-task scores it frames as dual-use, 84.5% on CyberGym and 54.4% on ExploitBench.⁵ These too are the lab's own numbers. Nathan Lambert's read is that Z.ai's release figures usually hold, but that is analyst judgment, and the open weights are staged about two weeks out pending a safety review.⁶ A large coding jump came from post-training alone, without retraining the base.

Cell therapy

First in vivo CRISPR screen in human T cells finds two tumor brakes⁷

A Gladstone and UCSF team ran what they describe as the first in vivo genome-wide CRISPR screen in human T cells, published in Nature on 12 August. Screening nearly 20,000 genes inside live tumors, they identified the P2RY8–Gα13 signaling axis and the gene GNAS as negative regulators of T-cell infiltration and function.⁸ Knocking out both improved CAR-T control across several solid-tumor models. In a lung-cancer model, two-thirds of mice given the doubly edited cells were tumor-free, versus none on controls. The team replicated the effect with T cells from ovarian and melanoma patients, and saw no long-term side effects over six months.⁹ The result is preclinical, in mouse models and patient-derived cells, not a human clinical outcome. Solid tumors have resisted the CAR-T therapy that clears blood cancers. This work points to two specific genes whose removal lifts that resistance, and to an in vivo screen others can now reuse.