HAI Daily

Meta's Muse Wonder, Qwen's QwenGyre, a New OpenClaw and Microsoft's Maia 300

Meta launches its first frontier models, Muse Wonder and Wonder Full, Alibaba's Qwen team shows QwenGyre for training long-running agents, OpenClaw adds skill search, MCP apps and full backups, and Microsoft unveils the Maia 300 inference chip.

By H.AI Tools Team4 stories
Products in this issue: Meta AI, Claude, ChatGPT, Qwen Chat, OpenClaw, CrewAI, Coze
In this issue

The short version

  • Meta's Muse Wonder costs $4 per million input tokens and $20 per million output tokens through its API. Meta says Wonder Full leads on its own index, but it trails Claude Opus 5.5 in third-party scoring.
  • QwenGyre speeds up reinforcement learning for agents that run for hours by moving idle GPUs between running tasks and training.
  • OpenClaw's October release lets agents search installed skills on demand, open MCP apps beside the chat and back up to external storage or Cloudflare R2.
  • Microsoft's Maia 300 inference chip is due in 2027. Microsoft claims 40% more tokens per megawatt than leading GPU systems.

October 13 brought Meta into the frontier-model race, new infrastructure for training and running agents, and a custom chip from Microsoft aimed at the cost of AI answers. Each section explains what was announced and how to use it.

01

Meta's Muse Wonder: Meta's first frontier models

What Meta launched

Meta released Muse Wonder and Muse Wonder Full on October 12, its flagship models for coding, long-running agent tasks and professional work. On Meta's own composite intelligence index, Wonder Full leads Claude Opus 5.5 and GPT-6 Astra, and Meta reports it ahead on several tests of automation, long-conversation memory and software engineering. In third-party composite scoring, however, it still ranks below Claude Opus 5.5. The standard Wonder model is available through Meta's API at $4 per million input tokens and $20 per million output tokens; reports say Wonder Full is limited for now to paid Muse Code plans.

The launch comes about 14 months after Llama 4 disappointed. Since then Meta has reorganised its AI work, set up its superintelligence lab and shifted to offering its best models through its own products first.

Comparing a new frontier model fairly

Vendor indexes favour the vendor. When a new model claims to lead, test it on your own work before switching. Take five tasks you run often, such as a code change with tests, a long document summary and a data question, and run them through Wonder and your current model with identical prompts. Score correctness, how many corrections you needed and the cost per finished task, not per token.

At $4 and $20 per million tokens, Wonder costs about the same as other top models, so the decision comes down to quality on your tasks and how well it fits your tools. Meta AI is the consumer entry point; compare it with Claude and ChatGPT.

Get HAI Daily in your inbox

The day's AI news, what it changes, and which tools to try.

02

QwenGyre: training agents that run for hours

What the Qwen team showed

Alibaba's (阿里巴巴) Qwen (通义千问) team introduced QwenGyre on October 12, a reinforcement learning framework for agents whose tasks last hours and run to close to a million tokens. It divides GPUs into elastic units: when a task finishes, its freed unit moves straight to training without disturbing tasks still running. It records each model call's real context and organises branching histories into a tree to pick useful training samples, and even unfinished tasks can contribute feedback. It treats the agent's harness as a black box, so an unmodified tool can be used for training.

In tests on NL2RepoBench, where each run handled around 700,000 tokens, the team reports end-to-end training up to about 1.8 times faster than an asynchronous baseline on the same GPU budget. The paper links no public code.

What it means for people using agents

Faster training for long tasks means agents that stay on track for longer, such as building a whole repository from a specification. For users, the practical lesson is to design work that agents can complete in checkable stages. Break a long task into milestones with tests or acceptance criteria, and have the agent report at each one. That makes today's agents more reliable and will suit tomorrow's longer-running ones. Qwen is the place to watch for models trained this way.

03

OpenClaw's October release

What changed

The open-source agent OpenClaw released version 2026.10.1 with three main features. Agents can now search installed skills and read a skill's full instructions when needed, instead of loading the whole skill catalogue into every prompt. Compatible MCP apps can open beside the chat, with support for forms and editing workspace files. Full backups can go to external drives or Cloudflare R2, with schedules, automatic cleanup of old backups and restores into a fresh staging directory rather than over existing files. The release also improves Mac installation, Code Mode and automation tasks. Some users report that frequent updates cause compatibility problems.

Running a personal agent safely

Skill search keeps prompts shorter, which means lower costs and fewer confused agents. Backups matter more: an agent that manages your files or messages needs a recovery plan. Turn on scheduled backups to a location the agent cannot overwrite, and test a restore into the staging directory once, so you know it works before you need it.

Because updates arrive often, read the release notes before upgrading and wait a few days after a major release if your setup is working. Install only skills you trust, and review what each can access. Our OpenClaw guide covers setup, and our OpenClaw vs Coze (扣子) vs CrewAI comparison helps if you are choosing a platform; Coze and CrewAI are the alternatives it covers.

Not sure which tool fits your task?Browse the H.AI Tools directory by category, then compare a shortlist.

04

Microsoft's Maia 300 inference chip

What Microsoft announced

Microsoft unveiled Azure Maia 300, its next AI chip, designed mainly for inference, the work of answering requests. Production deployments start in 2027, with plans to scale to multiple gigawatts. Reports say the computing chiplets use TSMC's 2-nanometre-class process, paired with HBM4 memory. A Maia 300 rack holds 96 chips, twice the previous generation, and Microsoft says it delivers up to seven times the rack-level throughput of Maia 200 and about 40% more tokens per megawatt than leading GPU systems; it has not said which systems it compared against.

Why custom chips matter to customers

Cloud providers build their own chips to lower the cost of serving AI and depend less on Nvidia. If that works, customers eventually see lower prices or higher usage limits, especially for high-volume inference. When Maia-based services arrive, compare price and speed for your workload against GPU-based options, and check whether the models you need are supported.

What to try next

Put one change into practice

Test any new frontier model on five of your own tasks before switching. Break long agent tasks into checkable milestones. If you run OpenClaw, set up scheduled backups and test one restore.

Explore Chatbots and Code Generation to build your shortlist.

How we put this issue together

The H.AI Tools Team checks each story against the official sources linked in the text and removes claims we could not confirm. Product links lead to H.AI Tools directory profiles. Unless an issue says otherwise, product notes describe vendor documentation, not our own benchmark tests.

Go deeper