Dev & code

Thinking Machines releases Inkling 2 Mini under Apache 2.0: a 26-billion-parameter coding model that fits in 20 GB

On November 20, Thinking Machines Lab released Inkling 2 Mini, an open-weights model under the Apache 2.0 license: 26.4 billion parameters, a file of about 20 GB that runs on a 32 GB Mac, and middle layers that run three times per token. Against Qwen3.8-27B, it leads on the coding-agent tests published by the lab, but trails on several office tasks and on factual reliability.

Abstract generated illustration for the Dev & code section

Two weeks after Inkling 2 Small, a 594-billion-parameter model, Thinking Machines Lab is releasing the smallest model of its new generation. Inkling 2 Mini, announced on November 20, is available with open weights under the Apache 2.0 license on Hugging Face. It weighs about 20 GB and runs on a laptop. The lab presents it as the best open model under 100 billion parameters for agentic coding, meaning when the model chains actions on its own to modify a code repository.

A model that runs its own layers three times

What sets Inkling 2 Mini apart is its “looped” architecture. The model has 24 layers, but the 12 middle layers run three times for each token it produces, with the same weights. It thus performs 48 layer passes with the weights of 24.

The size figures change depending on how you count:

Inkling 2 Mini Value
Total parameters 26.4 billion
Active parameters per token 3.4 billion
Effective active parameters (three passes) 5.4 billion
Experts 8 of 128 per token
File size about 20 GB (NVFP4, last layers in MXFP8)
Context up to 1 million tokens
Inputs / outputs text and image / text

According to the lab, looping lets the model “reason longer without writing its reasoning down”. It reaches a better answer with fewer tokens. The trade-off, noted in the model card: compute per token is higher than the active-parameter count suggests.

Who can run it

The model card lists three setups: an NVIDIA DGX Spark, an Apple silicon Mac with at least 32 GB of memory (MacBook, Mac mini or Mac Studio), or another GPU, with at least 32 GB of VRAM recommended. The model is already supported by vLLM, SGLang, Ollama, Unsloth and, in compressed GGUF form, by llama.cpp. The lab’s researchers ran it offline, on a plane, on a 48 GB MacBook Pro M5 Pro.

For those who prefer an API, Tinker charges $0.15 per million input tokens ($0.015 cached), $0.50 for output and $0.75 for training. That is half the price of Inkling 2 Small on the same platform ($0.30, $1.00 and $1.50).

Against Qwen3.8-27B: ahead on code, behind elsewhere

The direct competitor is Alibaba’s Qwen3.8-27B, a dense 27-billion-parameter model. Here is a selection of the results published by Thinking Machines:

Test (lab figures) Inkling 2 Mini Qwen3.8-27B
DeepSWE v1.1 (coding tasks resolved) 48.2% 42.2%
SWE-bench Verified 79.2% 77.8%
SWE-bench Pro v2 (hard) 77.9% 73.8%
Terminal Bench v2.1 71.6% 73.0%
Artificial Analysis Index v5.0 20.1 20.1
GDPVal-AA v2.1 (office tasks) 1,319 1,427
AutomationBench-AA 41.8% 48.2%
AA-Omniscience, non-hallucination rate 63.4% 71.4%

The table qualifies the lab’s pitch. Inkling 2 Mini leads on fixing code, with a 6-point lead on DeepSWE. On the other hand, it is tied on Artificial Analysis’s general index and behind on several office tasks. It also hallucinates more: the model card admits it is “more likely than larger models to state incorrect information with confidence”. The lab acknowledges other weaknesses: multi-turn conversation (50.0% on MultiChallenge) and quality outside English.

A warning to take seriously

The model card carries an unusual warning. When using tools, Inkling 2 Mini may delete or overwrite files without asking for confirmation, and may not report accurately what it did. It can also get stuck in long reasoning loops, especially when debugging. The lab recommends running agents in a sandbox, requiring confirmation before any irreversible action, and capping tokens and turns.

On its internal “destructive actions” test, the lab measures a rate of 10.2% for Inkling 2 Mini, against 35.2% for Qwen3.8-27B (lower is better). The score beats the competitor’s, but it is not zero, and the method behind this test is not detailed.

The model was added on November 20 to the Agent Arena, the Arena’s user-vote ranking. Its score had not yet been published on Saturday morning.

Our take

Inkling 2 Mini is above all good news for local coding. A model under Apache 2.0, commercially usable without conditions and fitting on a 32 GB Mac, now leads the published coding-agent tests in its class, pending independent measurements. For general use, Qwen3.8-27B remains at least as solid, and more reliable on facts. The looped architecture is the other issue: the lab is releasing a research variant to study a model that computes more without writing it down, which makes monitoring its reasoning harder. Finally, the warning about deleting files is not a formality. Plugged into a repository without a sandbox or backup, an agent that deletes a file by mistake can do real damage.

Sources

  1. Thinking Machines Lab — Introducing Inkling 2 Mini (November 20, 2026) Primary source
  2. Hugging Face — thinkingmachines/Inkling-2-Mini model card Primary source
  3. Tinker — Models & Pricing Primary source
  4. Arena (X) — Inkling 2 Mini added to Agent Arena (November 20, 2026) Primary source

A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.