Inkling 2 Mini reaches laptops; Space adds mobile editing
Inkling 2 Mini brings looped reasoning to laptops, Tencent releases Hy4 weights, and ChatGPT Space expands to Plus users and mobile editing. Browser automation, design handoffs and local coding tools also receive updates.
METR examines task horizons and judgment, while MiniMax reports cryptanalysis within a research scope. Runway and OpenHands describe concrete workflows for agent-assisted software delivery.
01
Models and weights
2 stories
2026-11-20Thinking Machines Lab
Inkling 2 Mini brings looped reasoning to a laptop-sized model
Thinking Machines Lab released open weights for Inkling 2 Mini, with inference and customization also available through Tinker. The model has 26.4 billion total parameters and 3.4 billion active parameters, and runs its middle layers three times per token. Accounting for those repeated passes gives 5.4 billion weight activations per token, a different measure from the active parameter count.
Most weights are trained and deployed in NVFP4, with MXFP8 in the final four layers, for a total memory footprint of about 20GB. The team demonstrated an offline flight-simulator build on a 48GB M5 Pro MacBook Pro. It also released a variable-loop research model to examine how extra internal computation affects performance and chain-of-thought monitorability.
2026-11-20Tencent Hy
Tencent releases Hy4 weights alongside a system card and GGUF links
Tencent announced open weights for Hy4, its natively multimodal flagship mixture-of-experts model. It lists 770 billion total parameters, 49 billion active parameters and a one-million-token context window. The official post provides links to GitHub, Hugging Face weights, GGUF downloads and the license, giving developers entry points for obtaining the model and its deployment materials.
The release also includes a Hy4 System Card containing safety evaluations. The August release was Hy4 preview, with the same published parameter counts. The new Hy4 download links bring together model weights, deployment materials, licensing terms and safety evaluations.
02
Products and workflows
4 stories
2026-11-20OpenAI
ChatGPT Space welcomes Plus users and adds mobile editing
ChatGPT announced access to Space for Plus users and the ability to edit on mobile. People can collaborate on interactive pages, slides and sites. Space in the sidebar also brings pages, slides, images, files and uploads into a shared home, keeping project inputs and the resulting work together.
An accompanying official post describes writing, research, image creation and interactive charts within pages, plus turning ideas into slides. Pages and slides can update automatically from connected tools, for example to incorporate new feedback into a to-do list, maintain a project tracker or refresh a trip itinerary. Plus access and mobile editing are the new elements emphasized in this announcement.
2026-11-20Browser Use
Browser Use explains its move from 24 browser actions to one Python tool
In a November 20 technical retrospective, Browser Use explains browser-use 1.0, released on November 10. A persistent Python execution environment with raw Chrome DevTools Protocol access replaces the 24 browser actions in version 0.x. Agents can retain variables and write operations for situations that a fixed menu of actions does not cover.
The team reports gains of 14–33 percentage points for the same three models on BU Bench v2.1, with costs falling to roughly 1/2.7–1/5.2 of their previous levels. The evaluation ran 200 tasks and scored a frozen 180-task set, with one run per model. Other defaults also changed, so the results describe the entire release rather than an isolated effect of reducing tool count.
2026-11-19Conductor
Conductor 0.97.0 adds inline HTML artifacts to local workspaces
Conductor 0.97.0 adds artifacts to local workspaces: HTML pages generated by agents can appear inline in a conversation. Users can explicitly reference an artifact with an @ mention and create or disable templates in settings. Selecting text in a transcript also adds a quoted comment to the composer for a focused reply.
Local agents can now search existing chats, while text added through the keyboard shortcut retains tables, code blocks, lists and headings. Messages sent by one agent to another chat can arrive during its current turn rather than waiting for that turn to finish. The updates bring artifact display, history search and cross-chat coordination into local workspaces.
2026-11-20Figma
Figma demonstrates a shared specification workflow for coding agents
Figma published a workflow using Specs and the Figma agent to assemble development context. In its fictional shopping-app example, Pickup, an engineer uses the Hex connector to retrieve city-level usage data and combines product requirements, a FigJam diagram, annotated designs and interactive code layers on the canvas before generating an implementation plan.
The engineer checks design-system, testing, observability and rollout requirements, resolves comments with teammates, then passes the spec link to Codex. Figma MCP lets the coding agent read the plan and its associated design context. Specs is in open beta across all plans on web, desktop and mobile. The example illustrates a workflow rather than measured outcomes from a customer deployment.
03
Research and technology
3 stories
2026-11-20Google Research
Google details the Gemini architecture behind Duplex calling
Google Research detailed how Duplex evolved from task-specific calling models into a Gemini-based architecture. Early systems were trained for narrow jobs such as restaurant reservations; later versions combined generative models with deterministic state tracking. The current stack accepts free-text task descriptions and uses Gemini for understanding, planning and dialogue, without rebuilding a separate conversation model for each business category.
Specialized acoustic and conversational-cadence models help handle compressed phone lines, background noise and turn-taking. Google says calls begin with disclosure that an automated system is calling. Businesses can opt out in their Business Profile settings or ask during a call for future calls to stop. The disclosure adds implementation and merchant-control details to the previously launched Gemini Tasks calling feature.
2026-11-20METR
METR measures a roughly 3.6-month doubling in task time horizons
METR released capability measurements through Claude Opus 5.5 and GPT-6 Astra. On the combined TH1.1 and TH2 suites, the human task duration corresponding to a model's 50% success rate follows a trend of roughly doubling every 3.6 months. TH2 is harder, so a given model's time horizon on the combined suite is lower than on TH1.1 alone.
Models usually matched or exceeded the recruited skilled humans on METR's coding and optimization tasks, but showed weaker judgment than experts on tasks without clear empirical feedback. Agent swarms helped more on verifiable tasks with fast feedback; their effects on other tasks were unclear or negative. These measurements are not a direct estimate of how much work across all occupations can be automated.
2026-11-20MiniMax
MiniMax reports cryptanalysis results on research schemes and public challenges
MiniMax described automated cryptanalysis by its internal research model, MIM. The team reports structural weaknesses in historical NIST competition candidates, peer-reviewed constructions and public challenges, including theoretical attacks turned into executable code. It released technical reports, proof-of-concept code and curated, validated reasoning traces; MIM itself remains an internal research model.
Examples include a practical collision attack on Waterfall-256 and extending practical key recovery against ARADI from the previously attacked five rounds to eight rounds. MiniMax explicitly says the work concerns academic designs, past candidates and public challenge parameters, and does not affect deployed cryptographic standards. It says findings were disclosed to the respective authors in advance.
04
Engineering practice
2 stories
2026-11-20Runway
Runway reports a five-minute median wait for first approval in its agent workflow
Runway described progressively handing CI repair, review, approval, product testing and incident response to agents. The company says cloud agents drove 38% of its merged pull requests in October, up from 0.1% in June. In its core API repository, the median wait for a first approval fell from more than two hours to about five minutes between September and mid-November.
Before its coordinator went live, the team logged decisions in read-only mode and had people assess every proposed action during short trials. Merge checks apply to the exact commit, external actions must be safe to repeat, and people take over when policy requires it or a step stalls. The figures describe Runway's own workflow, rather than a controlled comparison across engineering teams.
2026-11-19OpenHands
OpenHands tracks vulnerability fixes until they disappear from shipped-image scans
OpenHands published its container-image vulnerability workflow: inventory deployed images, scan with both Trivy and Grype, then let an agent choose lower-risk fixes, open pull requests and track their progress. The agent keeps at most three of its own PRs open simultaneously. A work-log entry stays until the vulnerability actually disappears from the latest scan results.
The team says fixable critical vulnerabilities in its own built images fell from a peak of 87 to five since the process began last month, while fixable high-severity findings fell from 1,997 to 188. Major upgrades, vulnerabilities without released fixes and release decisions still need human judgment. Minor upgrades can be bundled, and PRs closed by people are not automatically reopened. The workflow is available as an automation template.
05
Industry developments
1 story
2026-11-20The Guardian
Anthropic falls on its second trading day as the technology sell-off continues
The Guardian reported on November 20 that Anthropic shares fell about 9% on their second trading day to $88.30, approaching the $83 offer price. The stock had gained 17% on its first day; the subsequent decline gave back most of that initial rise. The record IPO did not halt the concurrent sell-off in AI-linked shares.
The report said the Nasdaq Composite erased its gains for 2026 and closed about 5% below its level at the start of the year, while the Dow Jones Industrial Average and S&P 500 entered correction territory. The 5% figure is a year-to-date comparison, not the day's decline. The reporting connects market pressure with investor concerns about the scale of spending on AI infrastructure.