How to Feed Web Pages to Inkling 2 Mini
Feed a web page to Inkling 2 Mini by converting that page to clean Markdown, then pasting the Markdown into a local chat or loading the saved file. Do this when the wording of one URL matters and the model is running on your own machine, for example through Ollama. Web2MD does not connect to Inkling or Ollama. It only prepares the page.
Thinking Machines Lab facts below were checked on the announcement, the model card, and the Tinker page, and are marked verified on 2026-11-22.
Short answer
Convert the page to clean Markdown with Web2MD, save the URL and the capture date, then paste or load that file into Inkling 2 Mini running locally. The model card lists Ollama for Apple silicon. Web2MD does not launch the model.
What did Thinking Machines actually release?
Thinking Machines Lab released Inkling 2 Mini on 20 November 2026 with open weights, small enough to run on a laptop. The announcement calls it a Mixture-of-Experts model: 26.4B total parameters, 3.4B active, and 5.4B weight activations per token because the middle layers loop. The model card, verified on 2026-11-22, repeats those totals, says each extra pass adds about 1.0B effective active parameters, and licenses the weights under Apache 2.0.
The announcement says the footprint is ~20GB. The model card says the checkpoint is about 20GB, in mixed NVFP4 and MXFP8, and lists one NVIDIA DGX Spark, Apple silicon with 32GB of memory or more, and other GPUs at 32GB of VRAM or more. The offline demo is a MacBook Pro (M5 Pro, 48GB) in airplane mode.
The announcement does not print one context-window number. The model card does: a window of up to 1M tokens. The Tinker page lists Inkling 2 Mini at 26.4B total, 5.4B active, and 256K max context. Those are three statements. Benchmark labels such as GraphWalks BFS 256K are not a fourth.
| Source, verified on 2026-11-22 | What it says about size and context | | --- | --- | | Announcement | ~20GB footprint. 26.4B total, 3.4B active, 5.4B activations per token with looping. No single context-window figure. | | Model card | Checkpoint about 20GB. Context window of up to 1M tokens. Apple silicon at 32GB memory or more. | | Tinker model list | 26.4B total, 5.4B active, 256K max context. |
The Hacker News item for that announcement, item 50581260, showed a score of 1,371 on 2026-11-22. Scores move. The page still has to be handed over as text.
Why does clean Markdown matter more on a laptop?
A hosted model can hide a messy paste inside a large window. A local Inkling 2 Mini is already spending the machine on the checkpoint. The model card's floor is 32GB of memory on Apple silicon, or 32GB of VRAM on other GPUs, for a checkpoint of about 20GB. Navigation, a cookie bar, and a footer are extra tokens in whatever room is left. Those pages do not say how many tokens remain after the weights load, so this article does not invent that remainder.
Looping makes leftover text slower. The announcement says the two middle blocks run three times. The model card says compute per token is higher than the active-parameter count suggests, that minimum reasoning effort does not fully disable reasoning, and that looped models are new in local runtimes, so speed varies. A menu you did not mean to send is still decoded, then run through those middle layers again. Up to 1M tokens on the model card and 256K on the Tinker list cannot both be the window your process opened. Clean Markdown does not raise that ceiling. It spends the ceiling on the paragraph you opened.
The Web2MD homepage's own figures say about 65% fewer tokens than raw HTML, and 1,000+ installs. The public benchmark measures characters and structure on MDN and Wikipedia and does not report a token percentage.
| | A copied page | Clean Markdown of the same page | | --- | --- | --- | | What the local model reads | The article, plus menus, ads, cookie text, and sidebars | Title, headings, body, and a source line you added | | What it costs on a laptop | Extra prefill, repeated through the looped middle layers | The source you meant to send | | What you can check later | Chrome mixed in with the paragraph | A sentence under a heading, plus the URL and the day | | Use it when | You actually need the chrome | You need the article, one docs section, or a thread |
Cutting token cost by starting from webpage Markdown is the general version of this for other models. Here the limit is the laptop, not an API invoice.
How do you turn one page into a file the local model can read?
- Open the exact URL and wait until the text you need is on screen. Expand a collapsed section first. A converter keeps what has rendered.
- Turn that view into Markdown. For a public page, convert the webpage to clean Markdown in the browser, or use the free tier: three public URLs a day, converted on Web2MD's server, as llms.txt states. The URL to Markdown page is the same job starting from a link.
- Put the title, the URL, and the capture date in the first lines. Save one file per page. When the live page changes, the file does not.
- On a public URL, the CLI can write that file. The CLI page documents
npx web2md-cli https://example.com --meta -o article.md. The CLI fetches the URL. It does not start Inkling, and it does not carry a login the way the browser does. - Start Inkling 2 Mini in a runtime the model card names for your hardware. On Apple silicon that list is Ollama and Exo Labs. Paste the Markdown, or load the file if that client can read one. Neither official page prints an Ollama model tag. Use the tag that runtime publishes. This article does not invent one.
- Ask for one sentence and the heading above it, then ask what the file does not contain. If the heading is missing from the answer, the model is not using the file.
Web2MD pricing lists the free tier, Pro Monthly at $9 a month, the $15 30-day pass, Pro Yearly at $49.99 a year, and Lifetime at $149. Signed-in conversion is a Pro path. Pasting into Ollama is not a listed integration, on any plan.
Which runtime are you actually loading?
The announcement and the model card do not publish the same software list. Both were verified on 2026-11-22. Follow the page you are reading, then that runtime's own install notes.
| | Announcement | Model card | | --- | --- | --- | | Local inference it names | SGLang, vLLM, Ollama, Unsloth, NVIDIA Dynamo, TokenSpeed, and Llama App | Apple silicon: Ollama, Exo Labs. NVIDIA: SGLang, vLLM, Unsloth, NVIDIA Dynamo, Tokenspeed. GGUF via Unsloth and llama.cpp | | Fine-tuning it names | SkyRL, Miles, and NVIDIA AutoModel | Miles, SkyRL, and NVIDIA NeMo AutoModel | | Also offered | Inference and customization on Tinker, chat on the Tinker Playground, weights on Hugging Face | The same distribution paths, without an Ollama tag on the page |
Ollama, vLLM, and SGLang appear on both pages. Neither page shows an ollama run tag or a serve command. A guessed flag would be a fictional setup. The handoff stops at the file.
The model card says the model accepts text and images and generates text. This workflow is the text of a page. Web2MD produces Markdown, not an image prompt. Tinker's 256K label is a hosted listing. It is not a measurement of the GPU in the laptop.
What about a page you can see only while signed in?
The homepage says Agent Bridge connects an agent to the rendered page in your existing Chrome session and does not bypass authentication. The Agent Bridge page says the bridge returns Markdown from content that session can already render, after you sign in and complete MFA yourself. It does not log in, clear a CAPTCHA, or add a permission. Chrome and the extension have to be running. Setup registers MCP clients it detects, including Claude Code, Cursor, Claude Desktop, and Codex. Ollama is not one of them. Convert in the browser, save the file, and paste or load it.
The free tier runs on the server and covers public pages. It does not carry your browser session. The CLI page says the same split: public URLs are fetched directly, and a page that depends on your login belongs on the browser bridge.
The announcement says knowledge is mostly a function of total parameter count, and that looping does not store more facts. BrowseComp-style tool use is not included when you only paste a file. If the answer needs a second URL, convert that URL too.
Which Inkling 2 Mini facts should you recheck?
Runtimes and context labels move. These answers match the announcement, the model card, the Tinker page, and the Hacker News item read on 2026-11-22.
Can a local Inkling 2 Mini fetch a web page by itself?
No. Tool-use scores on the announcement are benchmark results. A local chat reads the text you supply unless you add tools.
What context length does Inkling 2 Mini publish?
The model card says up to 1M tokens. The Tinker page lists 256K max context. The announcement states neither. Verified on 2026-11-22.
Which local runtimes does Thinking Machines name?
The announcement names SGLang, vLLM, Ollama, Unsloth, NVIDIA Dynamo, TokenSpeed, and Llama App. The model card splits that by hardware and adds Exo Labs, llama.cpp, and a GGUF path. Neither page prints an Ollama tag.
Does Web2MD send Markdown to Ollama or Inkling 2 Mini?
No. Send to AI hands Markdown to Claude, ChatGPT, Gemini, Perplexity, Grok, or DeepSeek. You load the file into the local model yourself.
Does Web2MD bypass a login or a paywall?
No. Agent Bridge uses the Chrome session you already have. The free tier converts public pages on the server.
Are the homepage token and install figures an independent test?
No. About 65% fewer tokens, and 1,000+ installs, are the site's own figures. The benchmark does not report a token percentage.
Related Articles
Cloudflare Browser Run /snapshot vs a Clipper
How to Feed Web Pages to Gemini 4 Neon as Markdown
Your AI Agent Can't Read Anything You Log Into — Why, and Why Cookies Aren't the Fix
Most Read
last 30 days- #1Can Claude Read Reddit? No — Here's the One-Click Fix (2026)
- #2HTML vs Markdown for LLMs: I Wasted 67% of My Tokens for a Year
- #3Reducing Token Waste in ChatGPT and Claude: 7 Techniques That Cut Costs 72%
- #4Obsidian Web Clipper Official Plugin 2026: Complete Guide + When You Need More
- #5Reddit JSON API vs Scraping: The Honest 2026 Comparison for Developers
Latest Articles
- 2026-11-21Cloudflare Browser Run /snapshot vs a Clipper
- 2026-11-20How to Feed Web Pages to Gemini 4 Neon as Markdown
- 2026-11-19Your AI Agent Can't Read Anything You Log Into — Why, and Why Cookies Aren't the Fix
- 2026-11-17How to Back Up Claude Chats to Markdown, Then Delete Them Safely
- 2026-11-16Can Claude Scan Websites? What It Can Read, What It Misses, and a Workaround