Google says writing in blocks makes its experimental AI up to four times faster
Block-based text generation, faster decoding from Xiaomi, Claude safeguards and Apple's rebuilt Siri.

Google released DiffusionGemma, an experimental text model aimed at local editing and filling gaps in code. It refines blocks of 256 tokens, the small text units models process, together instead of generating strictly one token at a time. Google reports up to fourfold faster generation than its one-token-at-a-time approach, with lower output quality than standard Gemma 4.
The Apache-2.0 release runs within 18 GB of GPU memory when quantized, using reduced numerical precision for model weights. Google reports more than 700 tokens per second on an RTX 5090. The speed benefits concentrate on dedicated accelerators handling relatively few requests together; the gains diminish at high concurrency.
Top Stories
Xiaomi and TileRT report more than 1,000 generated tokens per second from a trillion-parameter model. Their MiMo announcement puts the model on one eight-GPU commodity node, at approximately ten times standard MiMo-V2.5-Pro generation speed.
The system combines four-bit quantization of expert modules, specialized compilation and kernels, and speculative decoding: a smaller model drafts tokens that the larger model verifies. That reduces sequential generation work.
The reported number measures decoding speed, rather than complete task latency or production throughput. Access began as an application-based trial, with limited inference capacity.
Anthropic launched two models, then changed its stated policy for handling AI research requests. Fable 5 and Mythos 5 share an underlying model with different safeguards.
Fable launched for general use; Mythos had restricted access. Anthropic reports stronger performance on long-running engineering, vision and scientific tasks.
WIRED reports that Anthropic reversed a policy for silently limiting Fable's effectiveness on frontier-model development. The company said affected requests would instead receive visible refusals or routing to a less capable model. Anthropic also said visible safeguards could catch more benign requests.
Apple introduced a rebuilt Siri that searches personal context and acts across apps. Developer testing began with the new operating-system betas. Apple describes conversational answers and screen awareness, with examples including finding information in messages or email and editing and sharing photos.
Apple's system combines on-device and Private Cloud Compute models with an orchestrator using the local Spotlight index and App Toolbox. Third-party apps can expose personal-context information through Spotlight integration. Access requires supported hardware, and initial availability differs by platform, language and region.
Top Links
- Persistent work environments for agents (openai.com): OpenAI agreed to acquire Ona, whose cloud execution technology supports work beyond a laptop session; the deal remained subject to closing conditions and regulatory approvals.
- Safety when agents interact (deepmind.google): Google DeepMind and partners announced a research funding call of up to $10 million covering collective agent behavior, test environments, identity and oversight.
- An old geometry problem meets AI (sciencenews.org): New reporting examines May's AI-generated counterexample to an 80-year-old Erdős conjecture, with mathematicians questioning whether the mathematical result demonstrates a comparable advance in AI.
The week in AI, once a week.