LATEST
AI Tools

Inkling 2 Mini Is Live: Laptop AI Coding for $0.15 a Million Tokens

Updated November 21, 2026
15 min read
Editor's Note: This article was human-reviewed for accuracy
Close-up of JavaScript code on a laptop screen

Quick Answer & Key Takeaways

  • • Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, a 26.4-billion-parameter open-weights model that runs on a laptop.
  • • Its middle layers loop three times per token, lifting agentic coding to 48.2% on DeepSWE v1.1 while the whole model fits in about 20GB.
  • • Tinker charges $0.15 per million input tokens and $0.50 per million output tokens, and OpenRouter lists it at $0.10 and $0.40 through Prime Intellect.

Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, and it is the most unusual model launch of the month. It is a 26.4-billion-parameter open-weights model that runs on a laptop, yet it tops its closest rival on the headline agentic coding benchmarks. The trick is looping: its middle layers run three times for every token, so a small model does the reasoning work of a much larger one. Inkling 2 Mini is the second model in the Inkling 2 family, and it answers a question the first one left open. Inkling 2 Small proved the architecture at 594 billion parameters. Mini asks whether the same idea works at a size you can carry in a backpack.

The short answer is that it mostly does, with clear limits. Mini ties Qwen3.8-27B on the Artificial Analysis Intelligence Index at 20.1, leads it on DeepSWE v1.1 by six points, and costs about one-sixth as much per task. It is also weaker than Qwen on factual recall, multi-turn chat, and several knowledge-work benchmarks. This article covers what Inkling 2 Mini is, what it costs on Tinker and OpenRouter, how to run it locally, and where it falls short.

What a looped transformer actually does

A normal transformer runs each token through its layers once, from the first layer to the last. Inkling 2 Mini has 24 layers, but it does not treat them equally. The first six layers run once, the middle 12 run three times with shared weights, and the last six run once. That adds up to 48 layer applications per token from 24 layers of stored weights. Each extra pass through the looped layers adds about 1.0 billion effective active parameters, so the model behaves like it has 5.4 billion active parameters per token even though only 3.4 billion are active at any moment.

The passes are not simple repeats. Every pass starts from the output of the first six layers, and each pass after the first also reads a gated message from the previous pass. Thinking Machines says the middle layers carry the more abstract reasoning, while the first and last layers mostly translate between embeddings and internal representations. In its tests, looping the middle 12 layers beat looping the last 12 on 27 of 28 evaluations. The compute-equivalent gain from looping the middle layers was 1.51 to 1.91 times, against 1.24 to 1.36 times for looping the last layers.

Looping also changes what the model can do without writing its reasoning down. A looped model does more computation per token internally before it produces a token, so it can solve longer tasks with no visible chain of thought. Thinking Machines measured this with a variable-loop research model released alongside Mini. Its no-chain-of-thought time horizon rose from 16 seconds at one pass to 37 seconds at three passes. That is a research finding, not a product feature, but it is the part of this release that safety researchers will study most closely.

The model is a mixture-of-experts design with 128 routed experts per layer, eight of which activate per token. Most of its weights were trained and deployed natively in 4-bit NVFP4 precision, with the last four layers in 8-bit MXFP8. The total checkpoint is about 20GB. It accepts text and image inputs, produces text, and supports a 1 million token context window. The license is Apache 2.0, which permits commercial use.

The benchmarks: where Mini wins, and where it loses

Thinking Machines published a large benchmark table with its announcement, comparing Mini against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B, and Nemotron 3.5 Super VL. All of these scores are lab-run, so treat them as the vendor's own measurements until independent evaluators confirm them. Artificial Analysis, an independent benchmarking company, separately placed Mini at 20 on its Intelligence Index, tied with Qwen3.8-27B.

BenchmarkInkling 2 MiniQwen3.8-27BBest of the rest
Artificial Analysis Index v5.020.120.113.6 (Nemotron 3.5 Super VL)
DeepSWE v1.148.2%42.2%24.7% (Nemotron 3.5 Super VL)
SWE-bench Pro v2 (Hard)77.9%73.8%44.3% (Muse Glimmer)
SWE-bench Verified79.2%77.8%76.0% (Muse Glimmer)
Terminal-Bench v2.171.6%73.0%60.7% (Nemotron 3.5 Super VL)
Vals Index v2.246.5%Not reported29.8% (Nemotron 3.5 Super VL)
BrowseComp79.0%61.6%78.9% (Nemotron 3.5 Super VL)
MCP Atlas82.7%77.3%75.5% (Muse Glimmer)

The pattern is consistent. Mini leads on agentic coding and tool use: DeepSWE, SWE-bench Pro, SWE-bench Verified, BrowseComp, DeepSearchQA at 85.3%, and MCP Atlas. It also leads on several vision and document tasks, including OmniDocBench at 92.8% and ScreenSpot Pro at 76.5%. Where it loses, it loses on knowledge and office work. Qwen3.8-27B is ahead on AA-Briefcase (1391 against 1263), GDPVal-AA (1427 against 1319), AutomationBench (48.2% against 41.8%), EnterpriseOps-Gym (44.2% against 42.0%), and MMMU Pro (76.3% against 74.9%).

The factuality gap is the most important weakness. On AA-Omniscience, Qwen's non-hallucination rate is 71.4% against Mini's 63.4%. Thinking Machines says this plainly in its own model card: looping improves reasoning but does not let the weights store more facts, because factual memory tracks total parameter count rather than pass count. Mini compensates by using search and tools well, which is why it leads on BrowseComp and DeepSearchQA while trailing on closed-book factual tests.

Two other limits deserve attention. Mini scores 50.0% on MultiChallenge, a multi-turn dialogue test, and 33.4% on ComplexConstraints v1.1, both behind Qwen3.8-27B. And on Terminal-Bench v4.0, the newer harness, Mini and Qwen tie at a low 5.6%, which suggests both models struggle with that benchmark's harder tasks. The headline numbers come from the easier Terminal-Bench v2.1 harness, so do not read 71.6% as the model's general terminal ability.

What it costs to use Inkling 2 Mini

There are three ways to pay for Mini: Thinking Machines' own Tinker platform, an API marketplace such as OpenRouter, or your own hardware, where the token price is zero after the machine is bought. The prices below are per million tokens, verified on November 21, 2026.

RouteInputCached inputOutputTraining
Tinker (64K training context)$0.15$0.015$0.50$0.75
Tinker (256K training context)$0.15$0.015$0.50$1.50
OpenRouter (Prime Intellect)$0.10$0.02$0.40Not offered

Tinker is Thinking Machines' fine-tuning and sampling platform, and its prices are the first-party rates. OpenRouter's listed rates come from its sole provider for this model, Prime Intellect, and they undercut Tinker by a third on input and a fifth on output. OpenRouter's own data shows customers actually pay even less on average, about $0.033 per million input tokens and $0.397 per million output tokens, because of caching and discounts. There is one catch on Tinker: its serverless inference beta, the option aimed at production traffic, currently covers Inkling and Inkling-Small only, not Mini. On Tinker, Mini is available for inference and fine-tuning through the standard platform and for chat in the Tinker Playground.

At November 21, 2026 market rates of about £0.76 and C$1.43 per $1, Tinker's $0.15 input rate is roughly £0.11 or C$0.22 per million tokens, and the $0.50 output rate is roughly £0.38 or C$0.72. A million input tokens plus a million output tokens costs $0.65 on Tinker, about £0.50 or C$0.93, or $0.50 on OpenRouter. A typical coding-agent session with a million input tokens and 200,000 output tokens costs $0.25 on Tinker. Fine-tuning is cheap in absolute terms: 10 million training tokens cost $7.50 at 64K context or $15.00 at 256K context.

Artificial Analysis puts the cost in task terms. Using Tinker's first-party pricing, Mini costs $0.59 and takes 9.0 minutes per Intelligence Index task, against $3.53 and 38.2 minutes for Qwen3.8-27B. That makes Mini about six times cheaper and four times faster per task than its closest rival on that index, though Mini uses about 38% more output tokens to reach the same score. For comparison, GPT-6 Luna at maximum effort scores 28 on the same index for $0.15 per task, so Mini is not the cheapest way to get a high score. It is the cheapest way to get this score from an open-weights model you can run yourself.

Running it on your own machine

Local use is the point of this model, and Thinking Machines lists the hardware bar clearly. The checkpoint is about 20GB, and the lab recommends a Mac with at least 32GB of memory, one NVIDIA DGX Spark, or another GPU setup with at least 32GB of VRAM. At launch, the lab supports Mini on any Mac system with 32GB of RAM or more through Ollama. Quantized GGUF builds are available through Unsloth and llama.cpp for other hardware.

The framework support arrived on day zero, which is unusual for a model this new. vLLM, SGLang, Hugging Face Transformers, TokenSpeed, NVIDIA Dynamo, and the Llama App all support Mini, and Ollama and Exo Labs cover Apple silicon. For fine-tuning, the lab lists SkyRL, Miles, and NVIDIA AutoModel. The variable-loop research variant is also served by vLLM, so researchers can experiment with pass counts without building their own serving stack.

Getting started takes three steps. First, download the weights from Hugging Face under the thinkingmachines/inkling-2-mini repository, or pull the GGUF build from Unsloth if you want a quantized version for llama.cpp. Second, load the model in your framework of choice: Ollama on a Mac, vLLM or SGLang on an NVIDIA machine. Third, point your coding agent at the local endpoint. The model supports tool calling and structured outputs, so it works with agent harnesses that expect a function-calling API. OpenRouter also lists the model under the slug thinkingmachines/inkling-2-mini for hosted use.

Thinking Machines' own demos show what local use looks like. Its researchers ran Mini offline on a MacBook Pro with an M5 Pro chip and 48GB of memory, in airplane mode, and built a playable flight simulator in a few minutes. In a second demo, the same model powered all 20 residents of a small pixel town called Thinkyville on one offline MacBook, with each villager planning its day, gossiping, and answering notes left on its door. These are company demos, not independent tests, but they show the intended use case: persistent local agents, not one-off chat.

Inkling 2 Mini vs Inkling 2 Small

Mini is the second model in the Inkling 2 family, and the comparison with its larger sibling shows exactly what you give up for laptop size.

Inkling 2 MiniInkling 2 Small
Total parameters26.4B594B
Active parameters per token3.4B (5.4B effective with looping)17B
Artificial Analysis Index20.131
DeepSWE v1.148.2%72.1%
Context window1M tokens1M tokens
Checkpoint sizeAbout 20GBNot published as a single local file
Runs on a laptopYes, 32GB Mac or DGX SparkNo, datacenter-class hardware
Tinker price (input / output)$0.15 / $0.50$0.30 / $1.00
LicenseApache 2.0Apache 2.0
Weights releasedNovember 20, 2026November 12, 2026

Small is the stronger model by a wide margin on hard coding tasks, and it is the right choice if you are paying per token for difficult agentic work. Mini is half the API price and runs on hardware you may already own, which makes it the right choice for local agents, fine-tuning experiments, and workloads where the model stays resident. Our Inkling 2 Small review covers the larger model in full, including its own pricing and safety notes.

The catches

Mini's model card is unusually honest about its limits, and they are worth reading before you trust it with real work. First, factual accuracy. Mini is less capable than Inkling 2 Small across most evaluations, with the largest gaps on factual recall and knowledge-heavy questions, and it is more likely than larger models to state incorrect information with confidence. If your workload is mostly questions the model must answer from memory, a larger model will serve you better.

Second, reasoning loops. In internal testing the model sometimes became stuck in long reasoning loops, or repeatedly re-checked its own work without making progress, particularly when debugging code. Thinking Machines recommends setting limits on output tokens and agent turns. This is a real cost risk on hosted APIs: a stuck agent burns output tokens at $0.50 per million until your limit stops it.

Third, agentic reliability. When using tools, Mini is less reliable than larger models at following detailed instructions and project conventions. The model card warns that it may take destructive actions, such as deleting or overwriting files, without confirmation, and may not report the outcome of its work accurately. Its internal misbehavior rate is 9.8%, the lowest of the five models in its own comparison table, but 9.8% is not zero. Run agentic workloads in a sandbox and require confirmation before destructive actions. That is the lab's own recommendation, not ours.

Fourth, the monitorability question. Looping lets the model do more computation without writing its reasoning down, which makes chain-of-thought monitoring less complete. The lab's own tests found the variable-loop model's reasoning traces shrank from 3,703 tokens at one pass to 1,661 tokens at five passes on GPQA Diamond questions, while staying legible. It also found that Mini has very little control over its chain of thought: when asked to follow an instruction about its own reasoning, it complied at most 0.1% of the time in the tested range. The lab's conclusion is that at this scale it sees no meaningful reduction in monitorability, and that a weaker model, Qwen3.6-35B-A3B, detected almost all side-task attempts in its LinuxArena test. Larger looped models may behave differently, and the lab says agentic monitorability will be important to track as these architectures scale.

Fifth, the practical limits. OpenRouter lists a single provider, Prime Intellect, so hosted availability depends on one company. OpenRouter also notes the model is not currently available through EU in-region routing. Tinker's serverless inference beta does not cover Mini. And the variable-loop research variant, TML-d24-varloop-research, was trained mostly at three passes and may behave unpredictably at other pass counts. It is a research artifact, not a second product.

Who should use it, and who should skip it

Mini fits developers who want a capable coding agent running on their own machine. If you have a 32GB Mac or a DGX Spark, you can run a model that scores 48.2% on DeepSWE and 79.2% on SWE-bench Verified without sending code to a cloud provider. It also fits teams that want a cheap fine-tuning base: Apache 2.0 weights plus $0.75 per million training tokens on Tinker make it inexpensive to specialize for a narrow task. Our DaisyRose-27B review covers another free 27-billion-parameter coding model if you want to compare local options, and our DGX Spark review breaks down the desktop hardware Mini is designed to run on.

Skip Mini if your workload is mostly factual question answering, long multi-turn conversations, or office automation. Qwen3.8-27B beats it on AA-Briefcase, GDPVal-AA, AutomationBench, and factuality, and it is the better choice for knowledge work. Skip it too if you need a model with a published serverless production tier: Tinker's serverless beta does not list Mini yet, and OpenRouter has one provider. Finally, do not give Mini unsupervised write access to anything you cannot restore. The destructive-action warning in its own model card is the clearest signal in this release.

Common mistakes to avoid

Do not confuse total parameters with active parameters. Mini has 26.4 billion total parameters, but only 3.4 billion activate per token, and 5.4 billion effective activations once looping is counted. Quoting 26.4 billion as the model's working size overstates it by nearly eight times.

Do not assume the cheapest hosted price is the first-party price. OpenRouter lists Mini at $0.10 per million input tokens and $0.40 per million output tokens through Prime Intellect, below Tinker's $0.15 and $0.50. Check both before you commit a workload, and remember that OpenRouter's average paid price is lower still because of caching.

Do not run the variable-loop research model in production. TML-d24-varloop-research exists so researchers can study looping, and its own card warns it may behave unpredictably at pass counts other than three. Use Inkling 2 Mini for real work and the research variant for experiments.

Do not let an agent run without token and turn limits. Mini's documented tendency to loop while debugging code turns into a bill on hosted APIs and a stalled session locally. Set a maximum output length and a maximum number of agent turns before you start a long task.

Frequently Asked Questions

What is Inkling 2 Mini?

Inkling 2 Mini is a 26.4-billion-parameter open-weights multimodal model from Thinking Machines Lab, released on November 20, 2026. It is a looped transformer: its middle 12 layers run three times per token, which gives it 5.4 billion effective active parameters from 3.4 billion active ones. It accepts text and image inputs, produces text, and supports a 1 million token context window under an Apache 2.0 license.

How much does Inkling 2 Mini cost?

On Tinker, Thinking Machines' own platform, Mini costs $0.15 per million input tokens, $0.015 for cached input, and $0.50 per million output tokens. Training costs $0.75 per million tokens at 64K context and $1.50 at 256K. On OpenRouter, Prime Intellect lists it at $0.10 per million input tokens and $0.40 per million output tokens. Running the weights locally costs nothing per token after hardware.

Can Inkling 2 Mini run on a laptop?

Yes. The checkpoint is about 20GB, and Thinking Machines supports it on Mac systems with at least 32GB of memory through Ollama, on one NVIDIA DGX Spark, and on other GPUs with at least 32GB of VRAM. Quantized GGUF builds from Unsloth and llama.cpp make it runnable on a wider range of hardware, though the lab notes that local speed varies by runtime.

How is Inkling 2 Mini different from Inkling 2 Small?

Mini is the small sibling: 26.4 billion total parameters against Small's 594 billion, and 3.4 billion active against 17 billion. Small scores higher, with 31 on the Artificial Analysis Index against Mini's 20.1, and 72.1% on DeepSWE against 48.2%. Mini costs half as much per token on Tinker and runs on a laptop, which Small cannot do.

Is Inkling 2 Mini free to use?

The weights are free under the Apache 2.0 license, which permits commercial use, and running them locally costs nothing per token. Hosted use is paid: $0.15 per million input tokens and $0.50 per million output tokens on Tinker, or $0.10 and $0.40 on OpenRouter through Prime Intellect. There is no free hosted tier listed for Mini.

Is it safe to let Inkling 2 Mini run coding agents unsupervised?

No, not without safeguards. Its own model card warns it may delete or overwrite files without confirmation and may not report its work accurately. Run agentic workloads in a sandbox, require confirmation before destructive actions, and set limits on output tokens and agent turns. Its internal misbehavior rate is 9.8%, the lowest in its comparison table, but that still leaves room for a costly mistake.

Conclusion

Inkling 2 Mini is the first looped transformer most developers can actually run, and its numbers hold up: leading agentic coding scores in its size class, a 20GB footprint, and hosted prices from $0.10 per million input tokens. Its weaknesses are equally clear, and the lab published them itself: weaker factual recall than Qwen3.8-27B, shaky multi-turn chat, and an agent that can delete files without asking. If you want a local coding agent and you sandbox it, Mini is the strongest open option at this size today. Try it on Tinker Playground or pull the weights from Hugging Face and run it on your own machine.

Sources

15 min read
35 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more