Long Context

AIAI-generated summary Researchers from Google and the University of Texas at Austin posted a preprint about a model named Muir that reads ten million tokens. The preprint describes how the 1.34-billion-parameter model keeps a working state of 25 megabytes of float32 numbers.

AIAI-generated summary TimeOmni-Live is a model evaluated using LiveTS-Suite, a benchmark designed for Continuous and Live Time Series Understanding. The benchmark contains 5,249 streams, 15,555 annotated response events and 2.2 billion timesteps organized into 18 tasks across six capability groups.

AIAI-generated summary Google announces long-context training support in its MaxText framework using four context-parallel strategies on Cloud TPU hardware. Google says Llama 3.1 8B achieves 70 percent Model FLOPs Utilization on TPU 5p for sequence lengths up to 4 million tokens.

AIAI-generated summary Paid Claude plans provide a standard context window size of 200K tokens for chatting with Claude or using Claude Code and Claude Cowork. A larger 1M context window is available by selecting specific models through the command line when usage credits are enabled.

AIAI-generated summary Developer flaviocopes.com explains how coding agents and language models work using tools like Ollama, Claude Code, and Gemma. The guide examines tokenization, context windows, and why models can generate incorrect information with high confidence.

AIAI-generated summary Tushaar Gangavarapu, Gil Tabak, and Clemens JS Schaefer posted Muir on arXiv on November 5th 2026 as a 1.34-billion-parameter model that reads ten million tokens from a fixed working state. The authors report that training throughput matched the gated-delta baseline and ran about 2.7 times faster than the Transformer at 32,000 tokens.

AIAI-generated summary Amazon Web Services, Inc. announces that Kimi K3.1 Pro from Moonshot AI is now available on Amazon Bedrock. Moonshot AI states that Kimi K3.1 Pro is a natively multimodal agentic model featuring a 1-million-token context window.

AIAI-generated summary MiniMax announced that the M3.1 Flash model is available on FireworksAI. MiniMax says representatives from MiniMax and FireworksAI will discuss building the model for long-running agents at an upcoming event in San Francisco.

AIAI-generated summary Thinking Machines released Inkling 2 Small, a 594B-total and 17B-active multimodal MoE trained in 68 hours on 12T tokens. The model features a 1M context window and is available on Tinker with Apache 2.0 weights promised next week.

AIAI-generated summary Baidu has released the ERNIE 5.5 Flash Base open-source foundation model featuring a mixture-of-experts architecture with 280 billion parameters. According to Baidu, the model competes with similarly sized open foundation models in reasoning, code, and long-context tasks.

AIAI-generated summary AMD announced day-zero support and optimization for Kimi K3.1 Pro on AMD Instinct MI355X GPUs using the MXFP4 checkpoint and DSpark speculative decoding. AMD says the implementation combines memory-efficient sparse attention with decode context parallelism for long-context serving and low-latency interactive workloads.

AIAI-generated summary Hugging introduced StateBench, a benchmark designed to evaluate how language models track entity state changes across long documents. The publisher says tests showed recurrent and hybrid models generalize much farther than transformers when trained directly on these tasks.

AIAI-generated summary Moonshot AI has released Kimi K3.1 Pro as an open-weight, natively multimodal agentic model featuring a 1-million-token context window. Moonshot AI says the model delivers substantial improvements in coding, 3D creation, financial research, factual accuracy, and other agentic capabilities.

AIAI-generated summary Anthropic has released a public beta for message threads in its Messages API to speed up long-running Claude Code sessions. Anthropic says the feature reduces median time to first token by approximately 35% for requests containing between 500,000 and 1 million tokens.

AIAI-generated summary Google previews Gemini 4 Carbon as its strongest frontier model with capabilities for cybersecurity and long-horizon engineering. Google says the model starts testing with a subset of cyber defenders and trusted partners through the Fairwind Program.

AIAI-generated summary Amazon Web Services, Inc. makes the new Anthropic Haiku 5.5 model available in the Kiro IDE and CLI for the AWS GovCloud (US) Regions. Amazon Web Services, Inc. says developers can use Haiku 5.5 to summarize and compact long context, classify and route requests, process text in bulk, and make small edits across files.

AIAI-generated summary Anthropic previews message threads in public beta on the Claude Platform to help make long-running agents faster by avoiding the need to resend full conversation histories. Anthropic says that for requests containing 500k to 1M tokens, median time to first token drops by about 35 percent when message threads are enabled.