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 · ActionableTG: @neurogen_news (Neurogen)→ original

Thinking Machines Releases Open-Weight Inkling 2 Mini with 26.4B Parameters for 32 GB Macs

Thinking Machines has released Inkling 2 Mini, an open model for code, search and tool use that understands both text and images. It has 26.4B parameters (the model's "knowledge capacity"), but only 3.4B are activated for each token (a fragment of a word), so any M-series Mac with 32 GB of memory can handle it. The middle layers run three times, so the model reasons more deeply and writes less out loud. For local agents, this is a cloud-free option.

AI-processed from TG: @neurogen_news (Neurogen); edited by Hamidun News
Thinking Machines Releases Open-Weight Inkling 2 Mini with 26.4B Parameters for 32 GB Macs
Source: TG: @neurogen_news (Neurogen). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @neurogen_news (Neurogen) ↗
Editor-in-Chief
12 of 24 layers run three times48 layer applications per token
Filesabout 25 GB in the compact version, 35 GB in 8-bit
Key metric #3On PC, the RTX 5090 with 32 GB of VRAM is confirmed

Want to run an AI coding assistant on your laptop, without the cloud and without API bills? Until now, you had to choose between a small model that quickly falls apart on complex tasks and a large one that needs a server. Thinking Machines has released Inkling 2 Mini as an open model and offers a compromise.

The "26.4B" in the news headline means 26.4 billion parameters, the model's total size, its "knowledge capacity," which sits on disk.

But this is a MoE (mixture of experts) architecture: for each token, a small fragment of text, only some of the experts are activated, 3.4 billion unique active parameters. As a result, the model behaves in speed and memory more like a small model, and it can run on any M-series Mac with 32 GB of unified memory.

How the model is built

Imagine a large library where, for each question, you open only a few relevant shelves instead of reading every book. That is how MoE works.

The second feature is the looped transformer. The model has 24 layers, and 12 of the middle ones are run three times in a row with the same weights. As a result, each token gets 48 layer applications, and counting the repeated passes, that comes to 5.4B weight activations per token. The authors explain the idea this way: the network gets to take more reasoning steps internally and spends fewer tokens explaining out loud. This matters for agents: every extra token means waiting time and context memory.

To fit consumer hardware, the experts in the first 20 layers are trained and shipped in NVFP4 format, and those in the last four in MXFP8. These are compact number-storage formats that take up less space than standard ones.

What it can do

The model accepts text and images and responds with text, covering code, document analysis and screenshot interpretation. Tool calling, search and long agentic tasks in an environment with connected tools are claimed. The context window reaches up to 1M tokens, meaning up to a million text fragments in a single request.

In the authors' demo, the model built a playable flight simulator in a few minutes on a MacBook Pro M5 Pro with 48 GB, and it was done right inside the airplane. The source also says the model beat Qwen3.8-27B on benchmarks, but no specific tests or figures are provided. It is worth checking the Thinking Machines blog post and the model page on Hugging Face.

What you need to run it

  • Mac: any M-series model with 32 GB of unified memory is the minimum. With headroom, 48–64 GB is better, leaving room for the system, other apps and long requests.
  • NVIDIA PC: running is confirmed on an RTX 5090 with 32 GB of VRAM. The ready-made recipe uses a context of 32–64 thousand tokens. This refers specifically to graphics card memory; regular RAM is counted separately.
  • RTX 3090, 4090 with 24 GB and other cards: the exact minimum for GGUF (the file format for local running) has not yet been established.
  • Disk: about 25 GB for the compact version or 35 GB for 8-bit, plus extra space for model files.

You can run it this way: on Mac via Ollama, on Windows and macOS via Llama App, on Linux via llama.cpp. For the RTX 5090 there is a variant with SGLang. The weights and the GGUF build are on Hugging Face, and you can try the model in the demo chat.

Limitations and the practical takeaway

There are several caveats. A million tokens of context will require additional memory, and on a 32 GB configuration such a volume is not guaranteed. The 32–64 thousand tokens confirmed for the RTX 5090 is a realistic benchmark. Nothing definitive can yet be said about owners of 24 GB cards, and the comparison with Qwen3.8-27B is given without details.

What developers should do: if you are building local agents or code assistants and you have a Mac with 32 GB or more, Inkling 2 Mini is worth testing on your own tasks. A good test is a long chain of steps with tool calls, since that is where the claimed token savings from repeated passes should show up. If you have a 24 GB graphics card, it is better to wait until the community establishes the minimum requirements for GGUF.

Story

Key Questions Answered

How the model is built?

Imagine a large library where, for each question, you open only a few relevant shelves instead of reading every book. That is how MoE works. The second feature is the looped transformer. The model has 24 layers, and 12 of the middle ones are run three times in a row with the same weights. As a result, each token gets 48 layer applications, and counting the repeated passes, that comes to 5.4B weight activations per token.

What it can do?

The model accepts text and images and responds with text, covering code, document analysis and screenshot interpretation. Tool calling, search and long agentic tasks in an environment with connected tools are claimed. The context window reaches up to 1M tokens, meaning up to a million text fragments in a single request. In the authors' demo, the model built a playable flight simulator in a few minutes on a MacBook Pro M5 Pro with 48 GB, and it was done right inside the airplane.

What you need to run it?

Mac: any M-series model with 32 GB of unified memory is the minimum. With headroom, 48–64 GB is better, leaving room for the system, other apps and long requests. NVIDIA PC: running is confirmed on an RTX 5090 with 32 GB of VRAM. The ready-made recipe uses a context of 32–64 thousand tokens. This refers specifically to graphics card memory; regular RAM is counted separately. RTX 3090, 4090 with 24 GB and other cards: the exact minimum for GGUF (the file format for local running) has not yet been established.

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…