Thinking Machines releases Inkling 2 Mini, a 26B open-weight looped model, plus 5 more updates
Version 1
In Brief
- Thinking Machines — Inkling 2 Mini, a 26B open-weight looped model for local use
- Anthropic — Claude Code v2.1.330 adds Enterprise Admin Console sign-in and built-in charts
- Perplexity — Perplexity Computer now works in Chromium browsers via extension
- AWS — SageMaker JumpStart adds a model leaderboard for quality, speed and cost
- AWS — Bedrock AgentCore runtime instances add EC2 G7 GPUs
- AWS — Amazon Textract Insights is generally available
Details
Thinking Machines — Inkling 2 Mini, a 26B open-weight looped model for local use
Thinking Machines Lab released Inkling 2 Mini, a 26B-parameter looped transformer designed to run on a laptop, with open weights on Hugging Face.
- Inkling 2 Mini is a looped transformer with 26B parameters; Thinking Machines says it scores 20.1 on the Artificial Analysis Intelligence Index.
- It has 24 layers: the middle 12 loop three times, while the first and last six run once.
- Hugging Face describes it as a multimodal Mixture-of-Experts with text and image input and a 1M-token context; the weights are on Hugging Face under the Apache 2.0 license.
- It runs on Macs with 32GB of RAM through Ollama and on NVIDIA DGX Spark, and is available on Tinker for inference and fine-tuning and in the Tinker Playground.
- Thinking Machines also released a variable-loop research model so others can study how looping affects chain-of-thought monitorability.
Anthropic — Claude Code v2.1.330 adds Enterprise Admin Console sign-in and built-in charts
Claude Code v2.1.330 lets users sign in to organizations managed in Anthropic's Enterprise Admin Console and adds easel, a built-in mod that draws charts.
- After `claude auth login`, Claude Code can sign in to organizations managed in Anthropic's Enterprise Admin Console and applies the organization's settings, cloud provider or gateway.
- New built-in mod easel draws charts when asked and under MCP metrics results, on Pro, Max and, where available, Team and Enterprise on macOS and Linux.
- Automatic charts can be turned off in `/config`.
- Self-hosted runners get `--start-dir` / `SELF_HOSTED_RUNNER_START_DIR` to start work in a repository subfolder, plus new timing metrics in `/metrics`.
Perplexity — Perplexity Computer now works in Chromium browsers via extension
Perplexity Computer can now work directly in Chromium-based browsers through the Perplexity extension.
- Perplexity Computer now works directly in Chromium browsers through the Perplexity extension.
- Users hand off a task from the extension's side panel.
- Computer then works across the user's browser tabs and tools while the user watches.
- The extension is available from the Chrome Web Store.
AWS — SageMaker JumpStart adds a model leaderboard for quality, speed and cost
Amazon SageMaker JumpStart now offers a model leaderboard in SageMaker Studio to compare foundation models side by side.
- The leaderboard compares foundation models on reasoning quality, throughput, latency and cost in SageMaker Studio.
- AWS evaluates every model on BIG-Bench Hard, GPQA Diamond and MMLU-Pro with the same setup, and measures time to first token, throughput and instance cost per deployment configuration.
- Users can filter by provider, license and modality and pick a use-case profile and optimization goal (balanced, cost, latency or throughput).
- At launch it covers more than 70 LLMs from providers including Meta, Google, Qwen, IBM and Arcee, at no additional charge.
AWS — Bedrock AgentCore runtime instances add EC2 G7 GPUs
Amazon Bedrock AgentCore runtime instances now support EC2 G7 instances with NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs.
- AgentCore runtime instances now support Amazon EC2 G7 instances, powered by NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs.
- G7 types from g7.2xlarge to g7.48xlarge are added to a capacity provider's allowed instance types.
- AgentCore installs the NVIDIA drivers, so standard container images can run GPU workloads such as 3D rendering or vision models, in sessions of up to 14 days.
- Available in US East (N. Virginia, Ohio), US West (Oregon), Europe (Frankfurt, London, Stockholm) and Asia Pacific (Mumbai, Sydney); managed compute is charged on top of EC2 costs.
AWS — Amazon Textract Insights is generally available
Amazon Textract Insights, a generative AI feature that splits, classifies and extracts data from documents in one request, is now generally available.
- Textract Insights is a new generative AI feature of Amazon Textract, generally available in US East (N. Virginia) and US West (Oregon).
- It splits, classifies and extracts structured data from multi-document files in a single asynchronous request, with no models to train.
- Fields are defined in a blueprint, a JSON schema described in natural language, or taken from AWS public blueprints for documents such as W-2s, invoices and bank statements.
- Every extracted value comes with a confidence score and is grounded to its location on the page.