# Thinking Machines' Inkling 2 Mini: 26.4B looped open-weight model ties Qwen3.8 27B on Artificial Analysis

> Thinking Machines Lab released Inkling 2 Mini, a ~20GB looped mixture-of-experts model meant to run on a laptop. Artificial Analysis scores it 20 on its Intelligence Index, level with Qwen3.8 27B, at a fraction of the cost per task — but most benchmarks are the lab's own.

Published Nov 21, 2026 by TechMeasured Staff · AI · Canonical: https://techmeasured.com/ai/thinking-machines-inkling-2-mini-26b-looped-open-weights-ties-qwen3-8-27b-what-it-needs

Thinking Machines Lab on November 20 released **Inkling 2 Mini**, an open-weight model with 26.4 billion total parameters that the lab says is built to run offline on a laptop. Weights are on Hugging Face as `thinkingmachines/Inkling-2-Mini`, and the model is also served on the lab's Tinker platform.

It matters because it is a far smaller sibling to the 594B-parameter [Inkling 2 Small](/ai/thinking-machines-inkling-2-small-594b-moe-apache-weights-next-week) released earlier this month. Independent testing by Artificial Analysis puts it level with Qwen3.8 27B on overall intelligence, while being far cheaper and faster per task.

## What's different: a looped transformer

Most small models just stack layers. Inkling 2 Mini stores 24 layers in four blocks of six. According to Thinking Machines, the first and last blocks run once, while the two middle blocks share weights and run three times, so each token passes through 48 layer applications while only 24 are stored. It is also a mixture-of-experts model, pruned to 128 experts per layer during training, with 8 routed per token according to Let's Data Science.

That makes "size" slippery:

| Figure | Value | Source |
|---|---|---|
| Total parameters | 26.4B | Thinking Machines |
| Active per token (single pass) | 3.4B | Thinking Machines |
| Weight activations per token, counting loops | 5.4B | Thinking Machines; Artificial Analysis lists 5.4B active |
| Checkpoint size | ~20GB | Thinking Machines |
| Context window | 1M tokens | Artificial Analysis; model card via Let's Data Science |
| License | Apache 2.0 | Artificial Analysis; Let's Data Science |

The lab says the model was trained and is deployed natively in NVFP4 for 20 layers and MXFP8 for the last four, and claims no conversion loss. Its own demos ran on a MacBook Pro with an M5 Pro chip and 48GB of memory. Secondary coverage from Unite.AI and Let's Data Science says Macs with 32GB or more can run it via Ollama, and that GGUF quantizations are available for llama.cpp — Thinking Machines itself gives no minimum RAM figure.

## What the independent numbers say

Artificial Analysis, which tested the model on Thinking Machines' endpoint, reports:

- **Intelligence Index v5.0.1:** 20, against a median of 10 for comparable models; Thinking Machines' table lists 20.1 for both Inkling 2 Mini and Qwen3.8 27B.
- **Speed:** 280 output tokens per second; 9.72 seconds to first answer token.
- **Price:** $0.15 per million input tokens and $0.50 per million output, with a 90% cache discount; $0.59 to run each Index task.
- **Verbosity:** 161 million output tokens to complete the Index, versus an 85.7M median — a real cost for local users, where tokens mean time.

In a post quoted by TestingCatalog, Artificial Analysis put Qwen3.8 27B at $3.53 and 38.2 minutes per task, calling Inkling roughly 6x cheaper and 4x faster.

## The lab's own benchmarks

Everything else is self-reported. Selected rows from Thinking Machines' table:

| Benchmark | Inkling 2 Mini | Qwen3.8-27B | Gemma 4 31B |
|---|---|---|---|
| SWE-bench Verified (bash only) | 79.2 | 77.8 | 46.1 |
| SWE-bench Pro v2 (Hard) | 77.9 | 73.8 | 14.2 |
| Terminal Bench v2.1 | 71.6 | 73.0 | 43.4 |
| BrowseComp (with compaction) | 79.0 | 61.6 | 37.0 |
| MMMU Pro | 74.9 | 76.3 | 73.4 |
| ARC-AGI v2 | 49.3 | 42.4 | — |

Thinking Machines is candid about weaknesses: it says the model is not the strongest small model for multi-turn chat or instructions with many constraints, and it trails Qwen3.8 27B on MultiChallenge and ComplexConstraints. Its knowledge-accuracy score (AA-Omniscience) is -8.6 versus -6.4 for Qwen, with a 63.4% non-hallucination rate against Qwen's 71.4%. Let's Data Science also notes that Hugging Face described its own Qwen comparison as informal tests.

The lab additionally published safety and monitorability results, including internal "misbehavior" and "destructive actions" evaluations where it scores far lower (better) than rivals. Those tests are internal and unverified, and the lab calls its monitorability work preliminary.

## Who should try it

If you run coding or research agents locally on a 32–48GB Mac or a single large GPU, this is worth a test against Qwen3.8 27B — especially for agentic search and repo-level coding, where the lab's numbers are strongest. Budget for long outputs, and treat the benchmark table as a claim until more independent evaluations land. If you mostly chat or need tight instruction-following, the lab itself suggests looking elsewhere.


## Sources

1. [Thinking Machines Lab — Introducing Inkling 2 Mini](https://thinkingmachines.ai/news/inkling-2-mini/)
2. [Artificial Analysis — Inkling 2 Mini: Intelligence, Performance & Price Analysis](https://artificialanalysis.ai/models/inkling-2-mini)
3. [Let's Data Science — Thinking Machines Lab releases Inkling 2 Mini, an open-weight looped MoE](https://letsdatascience.com/news/thinking-machines-lab-releases-inkling-2-mini-an-open-weight-170080f4)
4. [TestingCatalog — Thinking Machines releases Inkling 2 Mini open-weight model](https://www.testingcatalog.com/thinking-machines-releases-inkling-2-mini-looped-open-model/)
5. [Unite.AI — Thinking Machines Lab Releases Inkling 2 Mini for Local Devices](https://www.unite.ai/thinking-machines-lab-releases-inkling-2-mini-for-local-devices/)
