cd /news/artificial-intelligence/inkling-2-mini-26-4b-open-weights-3-… · home › topics › artificial-intelligence › article
[ARTICLE · art-173644] src=shattered.io ↗ pub= topic=artificial-intelligence verified=true sentiment=↑ positive

Inkling 2 Mini: 26.4B Open Weights, 3.4B Active

Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, an open-weight multimodal model with 26.4B total parameters and 3.4B active parameters under an Apache 2.0 license, with weights on Hugging Face. The model pairs a looped transformer with a mixture-of-experts design, accepts text and image inputs, lists a 1M-token context window, and is also available through the company's Tinker platform for inference and customization. Thinking Machines calls it "Our smallest model, a looped transformer that runs on a laptop and excels at agentic coding," though no price, download size or benchmark scores have been published.

key takeaway

Inkling 2 Mini ships Apache 2.0 weights on Hugging Face with 26.4B total and 3.4B active parameters, a 1M-token context window and text-plus-image input, so teams can self-host or run it via Tinker.

by read17 min views4 publishedNov 21, 2026
Inkling 2 Mini: 26.4B Open Weights, 3.4B Active
Image: Shattered (auto-discovered)

Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, and it shipped with open weights. The company describes the model as its smallest, a looped transformer that runs on a laptop and excels at agentic coding. TestingCatalog AI News and Quartz both covered the release, and the weights sit on Hugging Face under an Apache 2.0 license.

The numbers on the sheet are short but telling. Inkling 2 Mini carries 26.4B total parameters and 3.4B active parameters, it pairs a looped transformer with a mixture-of-experts design, and it lists a 1M-token context window. It accepts text and image inputs and writes text back. Developers can also reach it through Tinker, the company’s platform for inference and customization.

This analysis separates what Thinking Machines and the coverage confirm from what nobody has published yet. That second list is long: no price, no download size, no benchmark scores. If you track how open-weight models trade memory against compute, our earlier piece on 280B mixture-of-experts versus 27B dense models gives the framework used below. For the wider beat, see the AI and machine learning section.

Don't miss new tech stories on Google

Add Shattered once in the Google app and our stories show up in your news suggestions.

Add now

What Thinking Machines Released on November 20 #

The release is a model, not a product bundle. Thinking Machines Lab put Inkling 2 Mini out as an open-weight model, which means anyone can download the trained parameters and run them without asking permission. The company also made it available on Tinker, where developers can run inference and customize the model. Those two routes cover the main ways a team would want to use it: self-host from the weights, or work through the company’s own platform.

The date matters for planning. November 20, 2026 is the release day, so anything you read about Inkling 2 Mini before that date was speculation. Everything in this article rests on the announcement and on the coverage that followed within a day. Independent testing has barely started, and that gap shapes how much weight you should put on the claims so far.

The official product description is short enough to quote in full. Thinking Machines calls it “Our smallest model, a looped transformer that runs on a laptop and excels at agentic coding.” Three claims sit in that sentence. The model is small relative to the rest of the lineup, it runs on laptop hardware, and it is good at agentic coding. The first is a matter of record. The second and third are marketing claims until someone measures them.

The phrase smallest model also tells us something by omission. A company only calls one model its smallest when larger ones exist or are planned. The sources cited here name no sibling models, so treat any talk of a full Inkling 2 lineup as unconfirmed until Thinking Machines publishes it.

Inkling 2 Mini Specs at a Glance #

The table below lists every spec the announcement and coverage confirm. Nothing here comes from leaks or estimates. If a detail does not appear in the table, assume it is still unknown.

Attribute Confirmed detail
Developer Thinking Machines Lab
Model name Inkling 2 Mini
Release date November 20, 2026
Model type General-purpose multimodal, text and image in, text out
Architecture Looped transformer with mixture-of-experts
Total parameters 26.4B
Active parameters 3.4B
Context window 1M tokens
License Apache 2.0
Weights Hugging Face
Hosted access Tinker, for inference and customization

Eleven rows, and every one is attributable. The table also shows what the announcement leaves out. There is no row for price, because no confirmed price exists in the cited sources. There is no row for benchmark scores or download size either, for the same reason.

Why 26.4B Total and 3.4B Active Parameters Matter #

The gap between 26.4B and 3.4B is the most important spec in the release. Total parameters measure how much the model knows and how much storage the weights occupy. Active parameters measure how much computation the model performs for each token it generates. Dividing 3.4 by 26.4 gives about 12.9%, so roughly one in eight parameters does work on any single step.

The mixture-of-experts side

A mixture-of-experts model splits part of its network into many specialist blocks and routes each token to a few of them. The rest stay idle for that token. That design lets a model carry a large pool of knowledge while paying the compute bill of a much smaller one. For a company pitching laptop use, it is a sensible choice, because generation speed tracks active parameters while quality tends to track the larger total.

The catch sits on the memory side. The full 26.4B parameters still have to live somewhere, because the router can call any expert at any moment. That is why a mixture-of-experts model feels fast but still demands memory closer to its total size than its active size. Thinking Machines has not published a download size, so we cannot say how big the files are or what a laptop needs to hold them.

The looped transformer side

A looped transformer reuses the same layers more than once for a single pass instead of stacking a fresh set of layers for every step of depth. The idea trades repeated computation for fewer unique parameters. Combined with mixture-of-experts routing, it points to a design that tries to get depth of reasoning without a giant parameter count.

The sources confirm the architecture label and nothing more. They do not say how many times the layers loop, whether the loop count changes at inference time, or how looping interacts with the expert routing. Those details decide whether the design delivers on the laptop promise, and they will come out of the technical documentation and early community tests.

The 1M-Token Context Window #

A 1M-token context window puts Inkling 2 Mini in the same headline range as the largest hosted models. Google’s Gemini 4 Neon also lists 1M tokens of context, as our report on Gemini 4 Neon pricing and thinking levels covered. The difference is that Neon runs on Google’s servers, while Inkling 2 Mini ships as weights you can host yourself.

A long window sounds simple, but it hides real costs. Attention memory grows with the number of tokens held in context, so a million-token session can eat far more memory than the weights themselves. A laptop might load the model and still choke on a full-length prompt. The announcement lists the window size, and it does not say how much memory a full window needs or how well the model retrieves facts from the far end of a long prompt.

For agentic coding, the window matters in a specific way. An agent that reads a repository, runs tests, and revises files accumulates a long trail of tool output. A bigger window lets it keep that trail instead of summarizing and losing detail. Whether Inkling 2 Mini uses the window well is an empirical question, and no one has answered it yet with published numbers.

Apache 2.0: What the License Allows #

Apache 2.0 is one of the most permissive licenses in common use for open models. It allows commercial use, modification, and redistribution, and it carries an express patent grant. It asks you to keep license and notice text and to mark changes. It does not impose the usage caps or field-of-use limits that some other open-weight licenses attach.

That choice carries weight for businesses. Legal teams at larger companies often block models released under custom licenses, because those licenses can restrict competing products or require a separate agreement above a user threshold. A standard Apache 2.0 grant clears most of that review quickly. Teams can fine-tune the model, bundle it in a product, and ship it to customers without a call to Thinking Machines.

One caveat applies to every open-weight release. A permissive license on the weights does not tell you what went into the training data, and it does not reveal the training recipe. The announcement does not claim to open either. Developers who need full reproducibility should read the license file and model documentation directly and not assume that open weights means open everything.

Running It on a Laptop: What Is and Is Not Confirmed #

The laptop claim drives much of the early interest, and it deserves the closest scrutiny. Thinking Machines says the model runs on a laptop. The sources do not confirm a download size, supported hardware, quantization formats, or inference frameworks. That means no one outside the company has published which laptops qualify, how much memory they need, or how many tokens per second to expect.

Laptops vary enormously. A thin ultrabook with limited memory and a workstation-class machine with a large unified memory pool both count as laptops in marketing copy. Until the company or the community posts real measurements, read the claim as a design goal. The 3.4B active parameter count supports it on the compute side, but memory is the deciding factor and the sources are silent on it.

If you want to prepare a test bench, our guide to running a local LLM for private AI walks through the usual setup. Once community builds appear, the Ollama API tutorial shows how to script against a local model with curl and Python. Neither guide mentions Inkling 2 Mini, so check that your runtime supports the architecture before you download anything. A looped design with expert routing may need updated tooling.

The Agentic Coding Claim #

Thinking Machines says Inkling 2 Mini excels at agentic coding. The company offers no benchmark scores in the cited sources, so the claim stands on its own for now. Agentic coding means the model does more than complete a snippet. It plans a change, calls tools, reads results, edits several files, and loops until a test passes.

That workload punishes small models in predictable ways. Agents make many sequential calls, so a slow model compounds delay across a task. Agents also fail when one bad tool call derails the plan. A model with 3.4B active parameters will be fast per step, which helps with the first problem. Reliability across long task chains is the open question, and it depends on training choices the announcement does not detail.

Readers who compare coding agents will find useful context in our look at Gemini CLI against Claude Code on token pricing. Hosted agents charge per token, and a local open-weight model changes that equation to a hardware cost. If Inkling 2 Mini holds up in real repositories, teams could move routine agent runs off paid APIs. That is an if, and independent tests will decide it.

Two Ways In: Hugging Face and Tinker #

The release offers two distribution paths, and they suit different teams. Hugging Face hosts the weights, so anyone can download them and load them into their own stack. Tinker offers inference and customization on Thinking Machines’ side of the fence. The first path gives control and privacy. The second gives convenience and, presumably, a managed route to fine-tuning.

Pricing for Tinker access to Inkling 2 Mini is not in the cited sources. That is worth stating plainly, because cost is the first question any team will ask. The self-hosted path has a different cost structure. You pay for hardware and electricity, and the license costs nothing. A team that already owns suitable machines can start testing the day the weights land.

Customization deserves a note of its own. Open weights let teams train adapters on private code and documents, then keep the result in-house. Our walkthrough on merging a LoRA adapter into a GGUF file covers that workflow for other models. Whether Inkling 2 Mini will ship in GGUF form is unconfirmed, since the sources name no quantization formats.

How Inkling 2 Mini Fits Among Open-Weight Rivals #

A fair competitive comparison needs published benchmarks, and none exist for Inkling 2 Mini. So we compare on design choices instead, using only what the sheet confirms. The model sits in the small, efficient corner of the open-weight field. Its 26.4B total size falls well below the largest open mixture-of-experts releases, and its 3.4B active count is small even next to mid-sized dense models.

The open-weight field is crowded and moving fast. Large players keep investing in it, a trend our report on Alibaba training its next Qwen on in-house chips followed. Another permissively licensed option widens the choice for buyers who weigh where their models come from. That question already draws attention from policy groups, as shown by the ASPI and ACSC guidance on open-weight provenance.

Where Inkling 2 Mini stands out on paper is the combination. Few small open models pair multimodal input, a 1M-token window, an Apache 2.0 license, and a looped design. Each trait exists elsewhere. The bundle is the pitch. Whether the bundle beats single-purpose rivals on coding quality per watt is something only tests can settle.

Historical Context: Small Open Models Keep Climbing #

Open-weight models used to trail hosted frontier systems by a wide margin. Over the past two years that gap narrowed, and the efficiency tricks behind it moved from papers into shipping products. Mixture-of-experts routing was an early example. Long context windows followed. Looped designs, which reuse computation instead of adding parameters, are the newer entry in that line of ideas.

Thinking Machines has pushed on a related theme with Tinker, a platform built around customizing models. Releasing Inkling 2 Mini as open weights fits that strategy. A small model that teams can adapt to their own data gives Tinker customers a reason to stay, and it gives the wider community a base to build on. The sources cited here do not describe the company’s wider roadmap, so we leave it there.

The pattern across the industry is consistent. Labs release a small, permissive model to win developer attention, then sell hosted services, larger models, or tooling on top. Inkling 2 Mini looks like a move from that playbook. It does not need to beat the best hosted model to succeed. It needs to be good enough that developers pick it as a default for local work.

Market Impact: Who Gains and Who Feels Pressure #

The immediate winners are developers who want a capable local model with clean licensing. An Apache 2.0 release with a 1M-token window gives them a new option for private code work, where sending source to a third-party API is a problem. Startups that fine-tune small models for niche tasks also gain a new base to test.

Hosted coding assistants face quiet pressure. Every capable local model gives enterprise buyers a stronger negotiating position, because the alternative to paying per token becomes real. That pressure arrives in a market already busy with large AI financing and IPO news, from Anthropic’s $83 IPO pricing to heavy data center spending. The money flows toward the largest models, and small open ones keep eroding the low end.

Hardware vendors benefit too. A model built to run on a laptop lifts demand for machines with large unified memory or strong integrated graphics. We cannot say which chips Inkling 2 Mini favors, because supported hardware is unconfirmed. Still, any serious local-model push strengthens the case for memory-rich consumer hardware.

What Remains Unconfirmed #

The gaps are as informative as the facts. The table below sorts the main questions into confirmed and unconfirmed. Use it as a checklist when new information arrives over the coming days.

Question Status What we know
Parameter counts Confirmed 26.4B total, 3.4B active
License Confirmed Apache 2.0
Context window Confirmed 1M tokens
Weights location Confirmed Hugging Face
Price Unconfirmed No price in cited sources
Download size Unconfirmed Not published in cited sources
Supported hardware Unconfirmed Laptop claim only
Benchmark scores Unconfirmed None in cited sources
Quantization formats Unconfirmed None named
Inference frameworks Unconfirmed None named
On-record executive quote Unconfirmed None found in cited sources

Four rows are confirmed and seven are not. That ratio is normal for launch day, and it argues for patience. Treat any viral benchmark chart or download size you see in the next 48 hours with suspicion until it traces back to Thinking Machines or to a reproducible third-party test.

Predictions: Five Things to Watch #

These predictions are our own analysis, not reported facts. Each one rests on the pattern of past open-weight releases and carries real uncertainty.

  • Community quantizations arrive within days. Open Apache 2.0 weights draw fast third-party conversions, and those will give the first real numbers on laptop memory use.
  • Early coding tests will split opinion. Small agentic models tend to shine on short tasks and stumble on long chains, so expect mixed first reports.
  • Framework support will lag the architecture. A looped, expert-routed design needs runtime updates, and some popular tools will take time to catch up.
  • Larger siblings will follow. The company calls this its smallest model, which suggests bigger versions are coming, though no source confirms a date.
  • Enterprise pilots will cite the license first. Clean Apache 2.0 terms remove a common blocker, so licensing, more than raw quality, may drive early adoption.

The first two predictions will resolve fastest. Watch community forums and Hugging Face discussion threads for memory measurements and tokens-per-second reports in the first week.

What Developers Should Do This Week #

Start with the model card on Hugging Face and read the license file in full. Confirm the file sizes before you plan hardware. Check your runtime’s release notes for support of the looped, mixture-of-experts architecture. Then test on a real task from your own codebase, not a public benchmark, because agentic performance depends heavily on your tools and repository layout.

Teams without suitable hardware can try the model on Tinker and avoid a download. Keep expectations modest and measure everything. Log task success rate, time to completion, and memory use for a few dozen runs, then compare against whatever model you use today. A small, honest evaluation beats any launch-day claim.

The takeaway: Inkling 2 Mini is a credible open-weight release with a clear design thesis and a permissive license. Its headline claims on laptop use and coding strength remain unproven. The numbers that matter most, memory footprint and agent success rates, will arrive from the community within days.

FAQ #

What is Inkling 2 Mini?

Inkling 2 Mini is an open-weight, general-purpose multimodal model from Thinking Machines Lab, released on November 20, 2026. It takes text and image inputs and generates text. The company calls it its smallest model, a looped transformer built for laptops and agentic coding.

How many parameters does Inkling 2 Mini have?

The model has 26.4B total parameters and 3.4B active parameters. It uses a mixture-of-experts design, so only a fraction of the network works on each token. That fraction is roughly 12.9% by our arithmetic.

What license does Inkling 2 Mini use?

The weights ship under Apache 2.0. That license allows commercial use, modification, and redistribution, and it includes a patent grant. Read the license text in the repository for the exact conditions.

Where can I download or try it?

Weights are available through Hugging Face. You can also use the model on Tinker for inference and customization. Pricing for the hosted route is not confirmed in the sources cited here.

Does Inkling 2 Mini really run on a laptop?

Thinking Machines says it does, but the sources confirm no download size, supported hardware, or quantization formats. The small active parameter count helps with speed. Memory needs depend on the total 26.4B parameters and on how much of the 1M-token context you use. Wait for community measurements before you buy hardware.

How good is it at coding?

The company says it excels at agentic coding. No benchmark scores appear in the cited sources, so the claim is unverified. Run your own tests on real repositories and track task success rates before you rely on it.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @thinking machines lab 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/inkling-2-mini-26-4b…] indexed:0 read:17min 2026-11-21 · —