Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableX: @MTSlive→ original

Thinking Machines Releases Inkling 2 Mini: A Laptop Model Six Times Cheaper Than Qwen3.8 27B

Thinking Machines has released Inkling 2 Mini, an open model that runs on a Mac with 32 GB of memory, so no cloud API is needed. Of its 24 layers, an average of 12 run three times for each token, so the model does more thinking internally and writes less. On the Artificial Analysis composite index it scores 20.1, the same as Qwen3.8 27B, but a task costs roughly six times less and finishes about four times faster. For local agents, it is a candidate worth testing.

AI-processed from X: @MTSlive; edited by Hamidun News
Thinking Machines Releases Inkling 2 Mini: A Laptop Model Six Times Cheaper Than Qwen3.8 27B
Source: X: @MTSlive. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceX: @MTSlive ↗
Editor-in-Chief
Apache 2.026.4B parameters, of which 3.4B are active
DeepSWE v1.1 (programming tasks)48.2%, the best among open models under 100B
Key metric #3The reasoning chain is shorter but still readable for inspection

A good model usually lives on someone else's servers: you pay for every request and send your code there. A model that fits on a laptop is typically too weak for serious agent work. Thinking Machines tried to remove that trade-off and released Inkling 2 Mini. It is an open model that runs on a Mac with 32 GB of unified memory. On the Artificial Analysis Intelligence Index, a composite ranking that compares models across reasoning and coding tests, it scores 20.1. Qwen3.8 27B scores the same. But according to the lab, Mini completes the same task about six times cheaper and roughly four times faster.

Mini is a response to the previous release. Some users found Inkling 2 Small too unwieldy at 594 billion parameters. In reply, the lab built a model about 22 times smaller, at 26.4 billion parameters.

How the model "thinks" in loops

A conventional neural network passes each token (a chunk of text) through a stack of layers once. Inkling 2 Mini is built differently. It has 24 layers, and an average of 12 of them run three times for each token. This architecture is called a looped transformer.

Think of a programmer who rereads a complex function in their head instead of writing out their reasoning on paper. The model does more computation internally and outputs less text. Fewer generated tokens mean less time and money per task. Also, of the 26.4 billion parameters, only 3.4 billion are active at once: only part of the network works on each token. That is why the model fits in a laptop's memory.

Quality and licensing

The lab reports the following figures:

  • 20.1 points on the Artificial Analysis Intelligence Index, on par with Qwen3.8 27B;
  • roughly 6x cheaper and about 4x faster per task at the same score;
  • 48.2% on DeepSWE v1.1, a software engineering benchmark. According to Thinking Machines, this is the best result among open models under 100 billion parameters.

The Apache 2.0 license allows commercial use, fine-tuning, and embedding in products without additional approvals. For teams that do not want to send code and data to an external API, this is the main practical advantage.

Can the model's "thoughts" be trusted?

Looped models carry a risk. If part of the reasoning happens inside the layers, the chain of thought the model writes in its answer may stop reflecting its actual reasoning. That matters for safety. From the announcement: "The model should not be free to write anything it likes in its chain of thought. The chain should show what the model actually does."

The lab tested monitorability, meaning how well the written reasoning reveals the model's behavior. The result: the chain became shorter but stayed readable. At Mini's scale, there is no noticeable degradation. Thinking Machines has also published a separate research model with a variable number of loops, so other researchers can study how loop depth affects transparency.

What to do with this

For now, these are the lab's own figures, and the announcement includes no independent measurements. A score of 20.1 places it at the level of an average open model, not a record. The claim of "best among open models under 100B" applies to a single benchmark, DeepSWE v1.1. Real speed and cost on your hardware and your tasks will have to be measured yourself.

The practical takeaway: if you are building a local coding agent or an internal tool where cost and privacy matter, add Inkling 2 Mini to your test suite alongside Qwen3.8 27B. Compare not just scores but time per complete task, since the 4x advantage is claimed per task, not per token. Also note that 32 GB of memory is the stated minimum for a Mac.

For the market, this is another signal: competition among open models is shifting from adding parameters toward using compute more intelligently per token.

Story

Key Questions Answered

Can the model's "thoughts" be trusted?

Looped models carry a risk. If part of the reasoning happens inside the layers, the chain of thought the model writes in its answer may stop reflecting its actual reasoning. That matters for safety. From the announcement: "The model should not be free to write anything it likes in its chain of thought. The chain should show what the model actually does." The lab tested monitorability, meaning how well the written reasoning reveals the model's behavior. The result: the chain became shorter but stayed readable.

How the model "thinks" in loops?

A conventional neural network passes each token (a chunk of text) through a stack of layers once. Inkling 2 Mini is built differently. It has 24 layers, and an average of 12 of them run three times for each token. This architecture is called a looped transformer. Think of a programmer who rereads a complex function in their head instead of writing out their reasoning on paper. The model does more computation internally and outputs less text. Fewer generated tokens mean less time and money per task.

What are the key details on Quality and licensing?

The lab reports the following figures: 20.1 points on the Artificial Analysis Intelligence Index, on par with Qwen3.8 27B; roughly 6x cheaper and about 4x faster per task at the same score; 48.2% on DeepSWE v1.1, a software engineering benchmark. According to Thinking Machines, this is the best result among open models under 100 billion parameters. The Apache 2.0 license allows commercial use, fine-tuning, and embedding in products without additional approvals.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…