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: @TeksEdge→ original

Open-Weight Inkling 2 Mini Matches Qwen3.8-27B on Independent Benchmark, Runs on 16 GB of Memory

Thinking Machines has released Inkling 2 Mini, an open-weight model that runs on a laptop with 16 GB of memory and, in the independent Artificial Analysis ranking, ties with Qwen3.8-27B. It runs 12 of its 24 middle layers three times, like an editor rereading a text: 48 passes over 24 stored layers. On top of that, only 5.4 of the 26.4B parameters are active for each word. For local assistants, this could mean skipping an expensive graphics card.

AI-processed from X: @TeksEdge; edited by Hamidun News
Open-Weight Inkling 2 Mini Matches Qwen3.8-27B on Independent Benchmark, Runs on 16 GB of Memory
Source: X: @TeksEdge. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceX: @TeksEdge ↗
Editor-in-Chief
Key metric #1Unsloth has already published GGUF versions for local deployment
Key metric #2A 1-million-token context on 16 GB is unrealistic
Key metric #3llama.cpp still requires a build with support for looped layers

A strong model usually means a lot of memory: the smarter the neural network, the more weights must be held in a graphics card or RAM, and running it on a laptop remains a dream. Thinking Machines, a US company, has released an open-weight model, Inkling 2 Mini, that tries to break this pattern. According to a post by reviewer @TeksEdge, on the Artificial Analysis Intelligence Index (an independent composite ranking that compares models across a set of tests), it ties with Qwen3.8-27B, a dense 27B-parameter model from the Qwen lineup. Inkling 2 Mini runs on a machine with 16 GB of memory.

How 24 Layers Become 48

Inkling 2 Mini has just 24 transformer layers. But instead of adding more, the authors loop the middle: the model runs 12 middle layers three times in a row for each token (a token is a chunk of a word). That makes 48 layer executions over 24 layers actually stored.

A everyday analogy: instead of three editors, one editor rereads the text three times, improving the draft each time. More computation is spent, but the "staff," that is, the set of weights, stays small. For local deployment this matters, because the bottleneck is usually memory, not the time of a single pass.

The second source of savings is the Mixture of Experts (MoE) architecture: the model consists of many "experts," and only some of them are activated for each token. Inkling 2 Mini has 26.4B parameters in total, but only 5.4B are active. That is roughly five times fewer than the dense Qwen3.8-27B, where all 27B are working. Fewer active weights means faster generation and lower load on the hardware.

What About the Numbers and Running It

The post lists Qwen3.8-27B at 20.1 on the Artificial Analysis Intelligence Index v5, while Inkling 2 Mini is given 48.2% and 77.9%, and the labels for these two values are mixed up in the post. But it does say something important: the Artificial Analysis index was evaluated independently, while the programming results come from benchmarks published by Thinking Machines itself. Check the exact scores on the Artificial Analysis website and the model card, and test the company's coding claims on your own tasks.

In practice, the picture for running it looks like this:

  • The weights are open, so the model can be run locally and fine-tuned.
  • Unsloth has released GGUF versions, including a compact quantization for 16 GB systems.
  • Macs with 32 GB or more are supported through Ollama.
  • Thinking Machines showed the model fully offline on a MacBook Pro building a playable flight simulator.

Limitations

First: the claimed 1-million-token context is unrealistic on 16 GB. The long-context cache will quickly consume memory, so on modest hardware you will have to work with much shorter windows.

Second: llama.cpp still requires a separate build with support for looped layers. A standard release cannot handle this architecture, and at first running it will take some fiddling with compilation or waiting until the support lands in the main branch.

Third: a comparison on a composite index does not guarantee superiority on your task. Agentic workflows, Russian-language work, and long tool-calling chains may show a different picture.

What to Do With This

If you are building local assistants, offline coding tools, or private pipelines, the release is interesting for two reasons. First, it demonstrates a working alternative to simply growing the parameter count: reusing layers buys more "thinking" without adding weights. Second, it is a strong American open model against the backdrop of Qwen and DeepSeek dominating the local segment with Chinese lineups.

Practical advice: if you have a laptop with 16–32 GB, download the Unsloth GGUF, build llama.cpp with the required support, and run the model alongside Qwen on your own tasks. Decide based on your own tests of speed, code quality, and stability in your context, not on the leaderboard.

Story

Key Questions Answered

How 24 Layers Become 48?

Inkling 2 Mini has just 24 transformer layers. But instead of adding more, the authors loop the middle: the model runs 12 middle layers three times in a row for each token (a token is a chunk of a word). That makes 48 layer executions over 24 layers actually stored. A everyday analogy: instead of three editors, one editor rereads the text three times, improving the draft each time. More computation is spent, but the "staff," that is, the set of weights, stays small.

What About the Numbers and Running It?

The post lists Qwen3.8-27B at 20.1 on the Artificial Analysis Intelligence Index v5, while Inkling 2 Mini is given 48.2% and 77.9%, and the labels for these two values are mixed up in the post. But it does say something important: the Artificial Analysis index was evaluated independently, while the programming results come from benchmarks published by Thinking Machines itself. Check the exact scores on the Artificial Analysis website and the model card, and test the company's coding claims on your own tasks.

What are the key details on Limitations?

First: the claimed 1-million-token context is unrealistic on 16 GB. The long-context cache will quickly consume memory, so on modest hardware you will have to work with much shorter windows. Second: llama.cpp still requires a separate build with support for looped layers. A standard release cannot handle this architecture, and at first running it will take some fiddling with compilation or waiting until the support lands in the main branch.

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…