Open-Weight Inkling 2 Mini Costs 6x Less Than Qwen3.8 27B: $0.59 vs $3.53 Per Task
Thinking Machines has released the open-weight Inkling 2 Mini. On one task from the Artificial Analysis index, it costs $0.59, while Qwen3.8 27B, which scores the same 20 points, costs $3.53, roughly six times more. The trick is repetition: the model runs 12 of its 24 layers three times each, much like an editor who rereads the same text three times, and this requires no new weights. That keeps it small enough to run on a MacBook, which makes it a new cost benchmark for local agents.
AI-processed from Artificial Analysis; edited by Hamidun News
Usually a "small" model means a trade-off: it is faster and cheaper, but noticeably dumber. Thinking Machines is trying to break that rule. Its new open-weight model, Inkling 2 Mini, solves one task from the independent Artificial Analysis benchmark for $0.59 in 9 minutes. Qwen3.8 27B, which scores the same 20 points on the Intelligence Index, spends $3.53 and 38 minutes on the same task. That is about 6 times more expensive and 4 times slower. The Artificial Analysis Intelligence Index is a composite score across reasoning, coding, and knowledge tasks: the higher the score, the stronger the model.
How does Inkling achieve such savings? It reruns some of its own layers several times instead of inflating its parameter count.
How the model works
Inkling 2 Mini is a mixture-of-experts model. It has 26 billion parameters in total, but only some of the experts are active for each token. It has 5.4 billion active parameters. That figure already includes three passes through the "looped" layers. Each text token uses 3.1 billion unique weights.
The idea of a looped transformer is simpler than it sounds. Imagine one person rereading a text three times instead of three different editors, noticing something new on each pass. In the model, 12 of the 24 layers run three times for every token. Compute goes up, but the memory needed for the weights does not. According to Artificial Analysis, that keeps the model small enough to run on a MacBook or a DGX Spark.
The license is Apache 2.0, so the model can be used freely in commercial products and modified. It accepts text and images, and its context window is 1 million tokens.
Numbers from the independent test
On the Intelligence Index, Inkling 2 Mini scored 20 points. Qwen3.8 27B (xhigh mode) and Nanbeige4.3-33B score the same. But Inkling has roughly 5 and 6 times fewer active parameters, respectively. On the intelligence-versus-size chart, the model sits on the Pareto frontier: no other model delivers more intelligence at the same size. This holds for both total and active parameter counts.
There is a downside. On one index task, Inkling produces about 152 thousand output tokens on average. Qwen3.8 27B produces about 110 thousand, so Inkling writes roughly 38% more. Nanbeige4.3-33B produces about 158 thousand. The repeated layer passes add compute but do not add tokens to the answer. The model simply reasons at considerable length.
Still, by the compute proxy metric (active parameters multiplied by token count), it uses about 2.8 times less than Qwen3.8 27B and 5 times less than Nanbeige4.3-33B.
Costs are calculated at Thinking Machines' own pricing: $0.15 per million input tokens, $0.50 per million output tokens, and $0.015 per million tokens found in the cache.
What to do with this
First, the caveats. A score of 20 is not flagship-level. The closed GPT-6 Luna (max) scores 28 and costs $0.15 per task, meaning it is both cheaper and stronger through the API. Inkling 2 Mini wins mainly among open models that can be deployed on your own infrastructure.
Who will find this useful:
- teams that need a local model for data privacy or offline operation;
- agent developers running long loops, where token price and latency matter;
- those who want to fine-tune a model, thanks to the Apache 2.0 license.
If you are currently running a 27–33B model on your own hardware, test Inkling 2 Mini on your own tasks. Pay close attention to response length. The extra 38% of tokens eats into part of the gain if you pay for output. The repeated-layer technique itself looks like a practical way to raise quality without increasing memory. It is worth watching whether other labs adopt it.
- Nov 21, 2026A viral GTA 6 Claude clip looks fake: half an hour of control burned through the 5-hour Pro limit
- Nov 21, 2026Claude Code Speeds Up Snowboard Kids 2 Decompilation: From 20% to 45% of Functions in Three Months
- Nov 21, 2026VibeCBS collects project nicknames in advance: the agent finds 2.8 times more documents
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
Key Questions Answered
How the model works?
Inkling 2 Mini is a mixture-of-experts model. It has 26 billion parameters in total, but only some of the experts are active for each token. It has 5.4 billion active parameters. That figure already includes three passes through the "looped" layers. Each text token uses 3.1 billion unique weights. The idea of a looped transformer is simpler than it sounds. Imagine one person rereading a text three times instead of three different editors, noticing something new on each pass.
What are the key details on Numbers from the independent test?
On the Intelligence Index, Inkling 2 Mini scored 20 points. Qwen3.8 27B (xhigh mode) and Nanbeige4.3-33B score the same. But Inkling has roughly 5 and 6 times fewer active parameters, respectively. On the intelligence-versus-size chart, the model sits on the Pareto frontier: no other model delivers more intelligence at the same size. This holds for both total and active parameter counts. There is a downside. On one index task, Inkling produces about 152 thousand output tokens on average.
What to do with this?
First, the caveats. A score of 20 is not flagship-level. The closed GPT-6 Luna (max) scores 28 and costs $0.15 per task, meaning it is both cheaper and stronger through the API. Inkling 2 Mini wins mainly among open models that can be deployed on your own infrastructure. Who will find this useful: teams that need a local model for data privacy or offline operation; agent developers running long loops, where token price and latency matter; those who want to fine-tune a model, thanks to the Apache 2.0 license.
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.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.