Meta has released Muse Wonder and Muse Wonder Full for personal and professional agent tasks, describing them as state-of-the-art models. Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers.
Meta has released Muse Wonder and Muse Wonder Full. Muse Wonder is available through the Meta Model API and Muse Code, and Muse Wonder Full is available to Muse Code subscribers. According to Meta, both models target personal and professional agent tasks and combine Active Memory with Active Focus to retain important context across conversations and sustain work toward complex goals. The company says they achieved state-of-the-art results on benchmarks covering personal and professional agents, memory, and coding.
Muse Wonder will be added to Muse experiences over time, rolling out first to subscribers, and will also support artifact generation and processes requiring more intelligence. Meta says both models were trained at its Prometheus data center and cover tasks in law, finance, health, and long-running software projects. Muse Wonder can also build a playable 3D kitchen game from customizable image inputs. Meta further explains that Muse Wonder Full performed above the level predicted by the Muse Spark 1.3 trend on unfamiliar tasks not targeted during training, attributing this to larger and stronger base models.
The X Subscription Launches, Integrating Grok, Cursor, and X Premium
要闻
x.ai has launched The X Subscription, bringing Grok, Cursor, and X Premium into one subscription. Grok and Cursor share a weekly included usage pool, while accounts, workspaces, and billing can be combined permanently after switching.
At checkout, the selected Cursor, SuperGrok, and paid X Premium subscriptions are cancelled, with automatic prorated refunds for unused time. A Cursor subscription paid through the Apple App Store cannot be cancelled or refunded by the service; the user must cancel it in iPhone or iPad Settings and request a refund from Apple. For Grok, Grok Bot, or X Premium purchases made through the Apple App Store, the service cancels the subscription and asks Apple to refund the unused time, while Apple decides whether to issue the refund. Users who purchased through Google Play must check their subscription status there after switching. A new billing cycle starts on the switching date. Users with multiple Cursor accounts can select one primary account and combine the others into it, but this option is not available for every account.
Tibo, the head of Codex, said the usage quota would be reset on day 8 of the 28-day plan, with the global reset to be completed before the end of October 12 in Pacific Standard Time.
Tibo, the head of Codex, said that usage quotas would be reset on day 8 of the 28-day plan and that the global reset would be completed before the end of October 12 in Pacific Standard Time. According to his explanation, the previously announced plan covers the next 28 days, during which the team will either deliver one clearly meaningful improvement each day to most Codex and Work users or fully reset their usage quotas.
智谱 Launches the Legal Model GLM-Legal and Opens Applications for Testing
模型发布
Zhipu has launched GLM-Legal, a legal model based on GLM-5.3-Flash for contract review, evidence analysis, and compliance workflows. API testing is open by application, and the capabilities will later be integrated into AutoClaw.
Zhipu has launched the legal model GLM-Legal, which is currently in testing; API access requires an application, and its capabilities are scheduled to be added to the desktop Agent platform AutoClaw. The model is based on GLM-5.3-Flash and supports workflows including contract review, evidence analysis, and compliance. In the official evaluation, GLM-Legal recorded a 19.2% full-pass rate on Harvey LAB, 12.5 percentage points higher than the base model. Zhipu said the model matched leading closed-source models on LegalBench and PLawBench with a substantially smaller parameter size.
OpenBMB Releases Two Models in the MiniCPM-V 4.7 Series
模型发布
OpenBMB released MiniCPM-V 4.7 on October 12, 2026, with 35B-A3B and 1B variants, and also introduced the 9B MiniCPM-o 4.5 for real-time processing of video, audio, speech, and text.
On October 12, 2026, OpenBMB released the 35B-A3B and 1B variants of MiniCPM-V 4.7 and updated MiniCPM-o 4.5. The 35B-A3B variant is the first MoE model in the MiniCPM-V series. According to the official description, Lookahead Expert Scheduling predicts experts that the next layer is likely to use across layers and prioritizes computation for experts that are already loaded, increasing generation speed by 5× on real devices. Weight quantization and an optimized storage layout also reduce storage use and improve the efficiency of scattered reads. According to the official description, this design reduces loading stalls when expert weights are stored on an SSD without sacrificing model capability.
MiniCPM-o 4.5 is a 9B end-to-end omnimodal model that continuously receives video and audio while generating speech and text at the same time, supporting real-time interaction across vision, hearing, and speech. The project was developed by THUNLP and ModelBest, and its code and model weights were released under Apache-2.0.
OpenCode Go Offers $80 Worth of V4.1 Flash Credits
开发生态
OpenCode has officially launched the third phase of “Operation Cheepseek,” giving OpenCode Go subscribers access to $80 worth of DeepSeek V4.1 Flash usage.
OpenCode announced that “Operation Cheepseek” has entered its third phase and that OpenCode Go subscriptions now include usage credits for DeepSeek V4.1 Flash. According to the company, the credits are worth $80 and apply to DeepSeek V4.1 Flash usage; the announcement does not specify other models or subscription benefits. The change described in the announcement is the addition of these credits to OpenCode Go, allowing subscribers to obtain the corresponding DeepSeek V4.1 Flash usage through the subscription.
阶跃 Step 5 Preview Launches WorkBuddy with Two Weeks of Exclusive Free Access
开发生态
Tencent has integrated Step 5 Preview into the WorkBuddy all-scenario AI workspace, offering exclusive free access for two weeks from October 12 to October 25, 2026.
Step 5 Preview is now available on Tencent’s WorkBuddy, with an exclusive free-access period scheduled for October 12–25, 2026. WorkBuddy is an AI office workspace developed by Tencent: users issue instructions in natural language, after which the system can break tasks down autonomously and call tools to move from requirement understanding to deliverable completion.
According to the official explanation, Step 5 Preview applies stronger reasoning and Agent capabilities to office workflows. It supports deep research, information retrieval, data analysis, document generation, and PPT generation, and can continue executing complex tasks involving multiple steps.
DeepSeek V4.1 Flash Is Free on the International Version of WorkBuddy Until October 31
开发生态
WorkBuddy announced that its international platform will continue offering free access to DeepSeek V4.1 Flash through October 31. The company said the extension was made in response to user requests.
WorkBuddy announced that DeepSeek V4.1 Flash will remain available at no cost on its international version through October 31. According to the company, the extension of free access was decided in response to requests from users. The available information does not specify whether the arrangement applies to other versions, the exact ending time, or any future pricing plan.
ZCode Launches an Invitation Campaign with Rewards of Up to 100 Days
开发生态
ZCode will run a referral campaign from October 12 to November 13, giving referrers a reward of 100 million tokens per day for up to 100 days for each successful invite. New users will receive a free allowance for five consecutive days.
ZCode will run a friend-referral campaign from October 12 through November 13, with the associated reward cards applying to GLM-5.3-Flash. For each successful referral, the referrer receives one Night “Cengceng” Card with an allowance of 100 million tokens per day, accumulated for up to 100 days. Referring two people grants a Weekend “Cengceng” Card containing 300 million tokens, while individual subscribers who reach 10 referrals in total receive a Weekly Reset Card.
Each invited new user receives a free allowance of 8 million tokens per day for five consecutive days, plus one Night “Cengceng” Card containing 100 million tokens. The campaign page also provides subscribers with randomized access to GLM-5.3-HighSpeed. Each session is free for 15 minutes, with a maximum speed of 500 tokens per second, and is limited to requests using GLM-5.3 while logged in.
OpenAI has added sandbox pre-warming to the Agents API, allowing developers to install dependencies, load files, and run setup commands before creating an Agent session; unattached pre-warmed sandboxes are free and expire after five minutes.
OpenAI’s official Agents API documentation states that developers can create an OpenAI-hosted sandbox, install packages, load files, and run setup commands before attaching the prepared environment to an Agent session. An environment is the workspace where an Agent runs tools, while a session stores the Agent configuration and conversation, so the two can be prepared independently. Container usage before attachment is free, and an unattached pre-warmed sandbox expires after five minutes; standard container pricing applies after it is attached to a session. An environment has the status pending while setup is running and becomes ready when setup finishes. Ready indicates that the workspace is prepared, not that it is connected to a session. A session can be created before setup finishes, but the Agent cannot access the workspace until setup is complete. Status changes can be obtained through a webhook or session events.
The create and attach examples require Python SDK 3.27.0, Node.js SDK 7.31.0, Java SDK 4.79.0, or Ruby SDK 0.102.0 or later, and environment creation uses POST /v1/agents/environments. A newly created shared environment supports up to 128 attached sessions, and multiple sessions can share the same workspace. An environment configuration can be saved as a template when the same packages and setup are needed repeatedly. For browser tasks, a managed browser and virtual display can be prepared before the session starts. GET /v1/agents/environments lists active environments created by the current identity in the current organization and project; results are newest first by default, with limit ranging from 1–100 and defaulting to 20. Using an Idempotency-Key during creation prevents duplicate environments when retrying after a timeout or lost response, and keys are retained for 24 hours.
Karpathy: The Perception Gap Around LLM Capabilities Is Widening
技术与洞察
Andrej Karpathy said that public understanding of LLM capabilities is splitting into a layered funnel: about 6 billion people have barely encountered LLMs, 1–2 billion occasionally use free chat products, about 20 million use frontier LLMs professionally, and about 5,000 can access internal frontier systems.
In this statement, Andrej Karpathy said that the gap in understanding LLM capabilities has developed beyond a divide between separate groups into a funnel structure that narrows by user population. His rough estimate puts the number of people who have barely encountered LLMs at about 6 billion, occasional users of free chat products at 1–2 billion, professional users of frontier LLMs in mathematics and coding at about 20 million, and people with access to internal frontier systems at about 5,000.
According to his explanation, professional users can already use prompts to have Agents complete large and complex coding projects that previously required weeks or months. Internal systems, meanwhile, take the form of thousands of Agents collaborating continuously for weeks. Karpathy attributed the gap to task scale and time span, differences in capabilities across fields, and differences in access among free, paid, and internal systems.
