Choose Your Edit Prediction Provider
Mercury Coder uses a diffusion architecture.
Mercury Coder uses a diffusion architecture.
Firecrawl v2.8.0 brings Parallel Agents, which run thousands of /agent queries at the same time, together with three new models: Spark 1 Fast, Spark 1 Mini and Spark 1 Pro. A new Firecrawl CLI covers the scrape, search, crawl and map commands. The Firecrawl Skill lets agents such as Claude Code, Codex and OpenCode use Firecrawl on their own, and self-hosted setups can now run linux/arm64 images.
Browser Use has released BU 2.0, which lifts accuracy on its benchmark from 74.7% with BU 1.0 to 83.3%, with an average task taking about 62 seconds against 58 seconds before. The company says it matches Claude Opus 4.5 on accuracy while running 40% faster. The bu-2-0 model is priced per million at $0.60, $0.06 and $3.50, compared with $0.20, $0.02 and $2.00 for bu-1-0.
Tabby v0.32.0 adds generic OAuth, indexing across several branches and the mistral/embedding API kind. Invitation links now carry the email address, which can no longer be edited when the link is opened.
From the official blog (lobehub.com), published . View original
This will guide you to select models and launch your chosen integration. No environment variables or config files needed.
FLUX.2 Klein handles readable text in images well, useful for UI mockups and designs with typography.
Existing applications using the Anthropic SDK can connect to Ollama by changing the base URL.
ERNIE-5.0 on LMArena is no longer labeled as “Preview” , indicating that the model has completed its preview phase.
By default, Codex will use the local gpt-oss:20b model.
I toned down some of the language on the landing page, as it had sometimes too much snake-oil-salesman-energy, should not oversell.
From the official blog (ernie.baidu.com), published . View original
From the official blog (ernie.baidu.com), published . View original
Browser Use has released BU-30B-A3B-Preview, its first open-weight model, with 30B total parameters and 3B active at inference. The company says it completes 200 tasks per $1, four times as cost-efficient as BU 1.0. Browser Use Cloud does not serve the model, so it has to be self-hosted, for instance with vLLM and connected through ChatOpenAI.
In highly technical manuals, many terms and phrases may appear semantically similar yet refer to very different mechanisms, conditions, or subsystems.
Your config doesn't just change colors, it rewrites the component code to match your setup.
ERNIE-5.0-Preview-1103 holds the top 20 in the most competitive Arena.
Firecrawl v2.7.0 offers ZDR search to enterprise customers, detects logos and colors more accurately in branding extraction, and takes screenshots faster and more precisely. A new minAge parameter waits until cached content reaches a set age before scraping again, and self-hosted deployments can choose their own crawl concurrency.
Browser Use Skills turn plain-text descriptions of an automation goal into API endpoints, with parameter and output schemas generated automatically. Skills can be created from the Dashboard or the API and run through the API with typed inputs.
Claude Code ships as a Bun executable to millions of users. If Bun breaks, Claude Code breaks.
I see Pocketbase, I upvote. Using it in a few production apps and it's been a very solid experience.
From the official blog (ernie.baidu.com), published . View original
Browser Use has launched a hosted HTTP MCP server at https://api.browser-use.com/mcp that any MCP client speaking HTTP, Claude Code among them, can use to control browser automation. Users who need a local stdio MCP server can still run the open-source uvx browser-use --mcp. Cloud projects can now add multiple team members who share billing and credits. Completed tasks can now get an LLM-as-a-judge evaluation, available in the library and the cloud.
Browser Use has opened a community workflow library at /templates, where templates can be cloned and run directly. Users can also contribute their own workflows.
Responding to strong community demand, we’ve significantly strengthened the model’s grounding performance with improved instruction-following capabilities, making visual grounding functions more accessible than ever.
From the official blog (ernie.baidu.com), published . View original
I discovered marimo after a shout-out from Pydantic creator Samuel Colvin .
marimo is, and always will be, free, open-source, and permissively licensed.
Keep in mind Raw CoT is meant for developers and safety practitioners. It’s not intended for exposure to general users or use cases outside of safety contexts.
Next.js works out of the box with the Bun Runtime on Vercel.
MiniMax M2 is now available on Ollama’s cloud. It’s a model built for coding and agentic workflows.
Converting a 2D visual document into a 1D text sequence inevitably discards spatial, contextual, and sometimes semantic relationships.
Browser Use can now sync a local Chrome profile to the cloud, so agents can use the authenticated sessions, cookies and preferences from your local browser. The sync runs through the profile.sh script once BROWSER_USE_API_KEY is set.
Folo v0.9.0 drops the cap on the number of views, lets you hide the All view and adds a right-click menu on views for hiding them or opening settings. The scrollbar now resets when you switch between entry details.
The test script and its readme are made available and can be customized for your own testing.
For users with more than 300GB of VRAM, qwen3-coder:480b is also available locally.
Browser Use open-source 0.9.0 introduces CodeAgent, which writes Python and JavaScript to interact with the browser through CDP using an LLM trained only for code generation. CodeAgent requires Browser Use's own ChatBrowserUse model and an API key from the cloud.
The newly released --mcp flag turns any notebook into a Model Context Protocol (MCP) server , exposing structured tools that let AI systems inspect, diagnose and reason about notebooks in a standard way.
Security logs from event sources were stored in Delta tables.