AI NewsOctober 13, 2026

MiniCPM-V 4.7 is an open-source model family that reads images, video and documents, in 35B-A3B and 1B sizes. A new scheduling trick lets the big one run on phones at 13+ tokens a second.
A 35-billion-parameter AI model normally needs a data center. MiniCPM-V 4.7 35B-A3B runs on an iPhone 17 Pro Max at 13+ tokens (word chunks) a second using about 3GB of memory, because only 3B of its 35B parameters work on each request.
Its Lookahead Expert Scheduling predicts which specialist parts are needed next and preloads them, like a waiter fetching the next course early, so phones can run capable vision AI without the cloud.
Key Capabilities:

• 5x speedup: Preloading experts speeds up an iPad Pro M4 by 5x, with byte-for-byte identical outputs.
• Two sizes: The 35B-A3B gives strong local performance, and a 1B version is easier to deploy.
• Open and broad: Free to download, with day-0 support for Apple, Intel, AMD, Arm and Rockchip chips.

QwenGyre is a training framework for AI agents whose tasks run for hours. It moves idle GPUs (AI chips) from finished tasks to learning, without stopping agents still working.
Teaching an AI agent to code a project can take hours per attempt, leaving expensive chips idle while slow attempts finish. QwenGyre fixes that like a restaurant moving idle cooks to prep work.
Using reinforcement learning (training by trial and reward), Qwen’s team finished 48 steps in 75.42 hours versus 134.55 for a standard setup. Faster training means better coding assistants reach people sooner.
Key Capabilities:

• Elastic scheduling: Frees chips from finished tasks for training while other agents keep running untouched.
• Better results: Qwen 3.8 2.4T’s coding test score rose from 52.5% to 58.5% over 48 steps.
• Existing tools: Trains agents through their current tool setups without rewriting them.

GLM-Legal is built on Z.ai’s GLM-5.3-Flash and tuned for contract review, evidence analysis and compliance. It scores 19.2% on Harvey’s strict LAB test, up 12.5 points from its base model.
Z.ai released GLM-Legal, a legal model built on GLM-5.3-Flash for contract review, evidence analysis and compliance. On Harvey LAB, a law-firm workflow test, it scored 19.2%, ahead of Kimi K2 (12.9%) and Claude Fable 5 (11.3%) in Z.ai’s chart.
That sounds low because Harvey grades all-pass: a task counts only if every step is right. Access is a beta by application, with no public price yet.
Key Capabilities:

• Legal focus: Trained on legal material and tens of thousands of legal task environments.
• Strict test: Harvey LAB covers 24 practice areas and counts a task only if every step passes.
• Apply for access: Request API access through Z.ai, or try it in the AutoClaw agent app.

Muse Wonder and Muse Wonder Full are built for personal and professional agent work, from coding to long-running tasks. Wonder is live on Meta’s API and Muse Code, and Full is limited to Muse Code subscribers.
Meta released two new frontier models, Muse Wonder and Muse Wonder Full, six weeks after Muse Spark 1.3. Artificial Analysis scores them 52 and 56 on its Intelligence Index, up from 48 for Muse Spark 1.3 at maximum effort.
Wonder costs $4 per million input and $20 per million output tokens on the API. Wonder Full comes only with Muse Code’s $50 Power and $200 Maximum plans.
Key Capabilities:

• Two models: Wonder is on Meta’s API and Muse Code, while Wonder Full is for Muse Code subscribers only.
• Memory and focus: Active Memory carries context across chats, and Active Focus keeps agents working through blockers.
• Coming to Muse: Wonder will soon power Meta’s personal Muse agent, starting with subscribers.

The X Subscription replaces separate Grok, Cursor and X Premium plans with five tiers from Free to Ultra. Paid plans start at $8 a month and top out at $200.
SpaceX Super Intelligence has merged Grok, Cursor and X Premium into The X Subscription, with five tiers from Free to Ultra. Paid plans share one usage pool across Cursor and Grok products, with at least twice the coding usage on Grok models as the equivalent Cursor plan.
The catch: third-party models inside Cursor now bill at API prices on the new plans. Existing subscribers can keep their old plan or switch.
Key Capabilities:

• Five tiers: Free, then Plus at $8, Premium at $30, Super at $100 and Ultra at $200 a month.
• Grok Bot included: Premium, Super and Ultra plans also include Grok Bot.
• Gradual rollout: Current subscribers can switch and link their Grok, Cursor and X accounts.