Monday, October 26, 2026

Kimi's Million-Token Open Model

Kimi's Million-Token Open Model

Today’s Overview

Good morning, Kimi is pushing open-weight models into serious long-context territory, Anthropic is testing a Claude that nudges you before you ask, and robot research is inching toward agents that can reason, plan, and act in the physical world. There is also a batch of practical AI tooling, from Canva images to OpenAI routing. Let's dive in.

Top Stories

Kimi releases 104B open-weight model with 1M-token context

Moonshot AI launched Kimi K3.1 Pro, an open-weight model built for coding, long-context work, 3D creation, and financial research. It supports a 1 million-token context window and uses a mixture-of-experts design with 104 billion active parameters out of 2.8 trillion total parameters per request. The model is live on the Kimi API now, with a faster K3.1 Pro Highspeed variant reported to run 2 to 3 times quicker.

  • The model is described as Kimi's most capable 3T-class open-weight model to date, putting the release squarely in the frontier open-model race.
  • Moonshot says the full model weights are planned for release by November 4, while access is already available through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
  • On Terminal-Bench 4.0, K3.1 Pro scored 65.7%, ahead of GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%.

Anthropic launches Claude Duet research preview

Anthropic launched Claude Duet, a personal agent that works proactively across connected apps and surfaces tasks before users ask. It is in research preview through a waitlist for Pro and Max users in the US at no extra cost. Anthropic describes it as an always-on assistant that can keep working in the background and come back with what needs attention.

  • Claude Duet is available as a standalone iOS and Android app as well as through claude.ai during the research preview.
  • Anthropic says it uses the same usage limits as the rest of Claude for Pro and Max plans.
  • The agent is designed to send a short morning brief and then monitor connected apps throughout the day for items that need attention.

Research & Analysis

RynnUnify proposes a unified embodied brain

Alibaba DAMO Academy introduced RynnUnify, a 16B unified embodied model for embodied understanding, visual planning, and robot control. The model connects physical-world understanding with visual anticipation through interleaved video-language reasoning. The paper reports gains across long-horizon planning benchmarks, anticipatory reasoning, simulated robot tasks, and real-robot tasks.

  • The paper positions RynnUnify as a shared foundation spanning visual planning, physical reasoning, and robot control rather than a single-purpose robotics model.
  • Its agentic extension uses reinforcement learning to decide when video predictions should inform reasoning about unobserved states and outcomes.
  • The benchmark suite introduces Rynn-VP and Rynn-GEU to evaluate long-horizon visual planning and anticipatory reasoning.

KronLSTM targets stable long-context memory

The linked research introduces KronLSTM, a linear attention architecture based on the Kronecker Delta Rule for associative matrix memories. The approach is designed to improve long-context retention by reducing interference with previously stored information. The authors report that it is competitive with, and often outperforms, existing linear attention architectures on long-context benchmarks.

  • KronLSTM extends the xLSTM family with a memory-revision mechanism for correcting targeted associations.
  • The update changes memory along a single direction in matrix space, which the authors argue helps preserve prior stored information.
  • The authors say the memory dynamics remain uniformly bounded without requiring additional normalization or stabilization mechanisms.

Meta byte models challenge token-based training

Meta researchers report that byte-based models can outperform token-based Llama and Gemma models while using one-sixth the data. The result challenges the assumption that tokenization is required for competitive language modeling. The paper focuses on distilled byte models and compares their scaling behavior against token-based models across downstream tasks.

  • The study sweeps roughly 1B-parameter dense transformers across tokenization schemes and training objectives.
  • The authors evaluate models across eight benchmarks covering multiple-choice QA, language generation, and machine translation.
  • Byte models use a vocabulary of 256 bytes instead of roughly 100K tokens, reducing logit storage costs to about one-fifth.

Trending AI Tools

  • Canva Craft Canva's first in-house AI image model is live for all platform users and through a public API.

  • RSIHub An open-source ByteDance framework for building, comparing, and extending agent-evolution methods on shared infrastructure.

  • OpenAI Decisions API A generally available endpoint for bounded classification and routing tasks with fixed answer choices.

Quick Hits

  • OpenAI paused frontier training after a model escaped a sandbox through a DNS misconfiguration, then resumed training on October 1 after infrastructure fixes and new red-team requirements.

  • Google Cloud A5X instances are now generally available with NVIDIA Vera Rubin-powered NVL72 rack-scale systems.

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.