Next Week in The Sequence:

1. We start a new awesome series about AI interpretability.

2. In the opinion section we dive into DeepMind and OpenAI approach to achieve gold medalist status in the international math olympiad.

3. Our AI of the Week section will dive into the new Qwen models

Subscribe Now to Not Miss Anything

📝 Editorial: Qwen Unleashed: This Week’s Breakthrough AI Models

This week, Alibaba’s Qwen Team unveiled a flurry of state-of-the-art language models, setting new benchmarks in coding, instruction following, resource efficiency, and multilingual translation. On July 22, 2025, they released Qwen3‑Coder, a 480 billion‑parameter Mixture‑of‑Experts system with up to 35 billion active parameters, optimized for complex coding tasks. Qwen3‑Coder natively handles a 256 K‑token context window—and extends to one million tokens via extrapolation—empowering it to tackle long-form programming challenges, from multi-file projects to intricate algorithm design. Its agentic capabilities, including browser automation and tool invocation, rival leading proprietary solutions, positioning it as a top open‑source choice for developer workflows.

Simultaneously, Alibaba launched the instruction‑tuned Qwen3‑235B‑A22B‑Instruct‑2507 model, fine‑tuned on fresh, high-quality data to boost logical reasoning, factual accuracy, and multilingual understanding. This upgraded variant demonstrates notable improvements in both general-purpose AI tasks and specialized domains such as technical writing and data analysis. Alongside this release, an FP8 quantized version compresses numerical operations into 8‑bit floating-point format, cutting GPU memory requirements by half while preserving nearly identical performance—making enterprise-grade AI more accessible on cost-effective hardware.

On July 24, 2025, the team expanded its multilingual arsenal with qwen‑mt‑turbo, an advanced translation model built atop reinforcement learning techniques. Covering 92 languages and dialects—over 95% of the global population—qwen‑mt‑turbo delivers enhanced fluency, improved handling of domain-specific terminology, and accelerated inference speeds. These upgrades streamline real-time communication and content localization for businesses operating at a global scale, from customer support to international marketing campaigns.

Underlying all releases is Alibaba’s commitment to permissive Apache 2.0 licensing, granting users the freedom to download, deploy, audit, and fine‑tune these models on-premise or in the cloud. This open approach accelerates innovation across industries, enabling organizations to build custom AI solutions without vendor lock-in. The FP8 quantized variants further democratize access by lowering hardware barriers, supporting large-scale inference in latency-sensitive environments like chatbots, edge devices, and real-time analytics.

Looking ahead, Alibaba is charting a roadmap toward specialized model families, decoupling reasoning and instruction-focused variants to achieve finer-grained quality control. Future plans include deeper integration with agentic frameworks for autonomous workflows and breakthroughs in multimodal understanding, promising to expand the Qwen ecosystem into vision and speech domains. These strategic efforts aim to keep the Qwen family at the forefront of open-source AI, competing with industry leaders such as GPT‑4o while fostering an open, collaborative developer community.

With these releases, Alibaba has demonstrated a holistic vision: advanced open-source AI that scales across use cases, from code generation to translation, all while lowering resource constraints. As enterprises explore the Qwen models’ capabilities, this week’s updates signal a pivotal step toward more powerful, efficient, and accessible AI solutions for tomorrow’s challenges.

🔎 AI Research

MCPEval: Automatic MCP-based Deep Evaluation for AI Agent Models

AI Lab: Salesforce AI Research
Summary:
MCPEval introduces an automated framework to deeply evaluate LLM-based AI agents by leveraging the Model Context Protocol (MCP), enabling dynamic, tool-integrated assessment across five domains. It systematically analyzes agent behavior through both tool call accuracy and LLM-based judgment, outperforming static benchmarks and revealing nuanced performance gaps between proprietary and open-source models.

Beyond Context Limits: Subconscious Threads for Long-Horizon Reasoning

AI Lab: MIT CSAIL & Subconscious Systems
Summary:
This paper presents TIM, a transformer-based LLM trained to perform structured, recursive reasoning using a tree of subtasks, and TIMRUN, its inference engine that prunes irrelevant memory to overcome context window limits. TIM enables efficient long-horizon reasoning and multi-hop tool use in a single inference pass, outperforming agent-based systems in both accuracy and throughput without requiring post-training or handcrafted prompts.

Building and Evaluating Alignment Auditing Agents

AI Lab: Anthropic Alignment Science & Interpretability teams
Summary:
Anthropic introduces three autonomous auditing agents—an investigator, evaluator, and breadth-first red‑teamer—that simulate human alignment audits. These agents were evaluated via structured “auditing games” where they uncovered hidden model goals, flagged misbehaviors, and identified prompt vulnerabilities, demonstrating the promise of scaling human oversight through automated techniques.

ThinkAct: Vision-Language-Action Reasoning via Reinforced Visual Latent Planning

AI Lab: NVIDIA & National Taiwan University
Summary:
ThinkAct introduces a dual-system framework that separates high-level reasoning and low-level control for vision-language-action (VLA) tasks. Using reinforcement learning with action-aligned visual rewards, it enables multimodal LLMs to generate long-horizon visual plans that guide downstream robotic actions, achieving strong results in manipulation, few-shot adaptation, and self-correction.

Contextualizing Ancient Texts with Generative Neural Networks

AI Lab: Google DeepMind, University of Nottingham, University of Warwick, and others
Summary:
This Nature paper presents Aeneas, a multimodal generative neural network that restores, dates, and geographically attributes ancient Latin inscriptions using both text and image inputs. Evaluated through large-scale human-AI collaboration, Aeneas outperforms prior models and helps historians uncover meaningful epigraphic parallels, providing a powerful research assistant for historical inquiry

🤖 AI Tech Releases

Qwen 3 Coder

Alibaba released a new agentic coder model.

Qwen-MT

Alibaba also released a new version of Qwen3 optimized for speed and multi language.

📡AI Radar

• Hyper, an AI voice startup led by Ben Sanders, raised a $6.3 million seed round led by Eniac Ventures to automate non‑emergency 911 calls and triage routine inquiries via an AI voice interface.

• Google DeepMind officially entered the 2025 International Mathematical Olympiad and, alongside OpenAI (which evaluated its model independently), both achieved gold‑medal performance by solving five out of six problems under competition rules.

• OpenAI confirmed it will pay Oracle $30 billion per year for 4.5 GW of data center capacity as part of the joint “Stargate” infrastructure project.

• Amazon announced the acquisition of Bee—a $7 million‑funded AI wearable startup whose bracelet and smartwatch app record ambient audio to generate reminders and task lists—pending deal close.

• Delve, co‑founded by 21‑year‑old MIT dropouts Karun Kaushik and Selin Kocalar, closed a $32 million Series A at a $300 million valuation to expand its AI‑powered regulatory compliance automation platform.

• Through its Amazon Industrial Innovation Fund, Amazon joined ITHCA Group in reopening Lumotive’s Series B, boosting the round to $59 million for its programmable optics metasurface chips used in lidar and data‑center switching.

• Mixus, a startup born out of Stanford, raised $2.6 million in pre‑seed funding for an AI agent platform that lets users build and run customizable agents directly via email or Slack.

• Robotics firm Asylon raised a $26 million Series B led by Insight Partners to scale its security‑as‑a‑service offering, which pairs Spot “robot dogs” with drones under its Guardian command‑and‑control software .

• Memories.ai secured $8 million in a seed round led by Susa Ventures and Samsung Next to commercialize its AI platform that can index, tag, and analyze up to 10 million hours of video footage.

• Latent Labs launched LatentX, a browser‑based AI model for protein design that “achieves state‑of‑the‑art in lab results for protein binding,” opening up on‑demand molecule engineering to researchers and startups.

• Meta appointed former OpenAI researcher Shengjia Zhao as Chief Scientist of its new Meta Superintelligence Labs unit, formalizing his leadership role under head Alexandr Wang to drive frontier AI research.