Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Inkling 2 Mini

Thinking Machines' Apache-licensed 26B looped model fits a 32GB Mac and matches Qwen3.8-27B on the lab's own index.

Reviewed by ReviewRamble editors · Updated

Visit Inkling 2 Mini Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

4.3/5

Inkling 2 Mini is the open model Thinking Machines owed after holding back Inkling 2 Small: a 20GB download that runs on a laptop and keeps pace with Qwen3.8-27B.

What is Inkling 2 Mini?

Inkling 2 Mini is the small, downloadable member of Thinking Machines Lab's second Inkling generation, released on 20 November 2026 under Apache 2.0. It is a mixture-of-experts model with 26.4 billion parameters and 3.4 billion active per token, routing each token to 8 of 128 experts. Its 24 layers are arranged so that the middle 12 run three times with shared weights, which the lab calls a looped transformer: each token passes through 48 layers while the checkpoint stores 24, for about 5.4 billion effective active parameters. It reads text and images, writes text, and supports up to 1 million tokens of context.

The official checkpoint is about 20GB in mixed NVFP4 and MXFP8 precision and is meant for one DGX Spark, an Apple silicon Mac with 32GB or more, or a GPU with at least 32GB of memory. ggml-org publishes GGUF builds for llama.cpp at 52.2GB (BF16), 27.8GB (Q8_0) and 17.6GB (IQ4_NL). Its larger sibling, Inkling 2 Small, is still closed.

How much does Inkling 2 Mini cost?

PlanPriceWhat you get
WeightsFreeApache 2.0, self-hosted
Tinker inference$0.15 / $0.50 per 1M tokensInput / output; cached input $0.015
Tinker training$0.75 (64K) or $1.50 (256K) per 1M tokensLoRA fine-tuning

Free weights under Apache 2.0. Tinker: $0.15 per 1M input tokens ($0.015 cached), $0.50 per 1M output; training $0.75 per 1M tokens at 64K context, $1.50 at 256K.

Compare this price with other models

Is Inkling 2 Mini worth it?

Inkling 2 Mini is the strongest reason yet to run a Thinking Machines model at home, though every score behind that claim is the lab's own. On its model card it ties Qwen3.8-27B on the Artificial Analysis Intelligence Index v5.0 at 20.1, ahead of Nemotron 3.5 Super VL (13.6), Muse Glimmer (6.3) and Gemma 4 31B (4.4). Hugging Face's launch post counts it first on 32 of the 48 benchmarks in that table, and its staff call it at least comparable with Qwen3.8-27B in their informal tests.

The loop is what makes the size work. Running the middle 12 layers three times gives the model 48 layers of computation from 24 layers of weights, so the checkpoint is about 20GB while each token gets roughly 5.4 billion parameters of work. The lab's table puts it at 79.2% on SWE-bench Verified with bash only, against 77.8% for Qwen3.8-27B, 77.9% on the hard SWE-bench Pro v2 split against 73.8%, and 82.7% on MCP Atlas against 77.3%. On vision it reaches 92.8% on OmniDocBench v1.6 and 76.5% on ScreenSpot Pro, where Qwen scores 91.5% and 57.6%.

It does not win everywhere. Qwen3.8-27B is ahead on Terminal Bench v2.1 (73.0% to 71.6%), AutomationBench-AA (48.2% to 41.8%) and MMMU Pro (76.3% to 74.9%), and both score 5.6% on Terminal Bench v4.0. Its non-hallucination rate on AA-Omniscience is 63.4%, below Qwen's 71.4%, and the card says it trails Inkling 2 Small most on factual recall.

The catch: the lab's own limitations section reads like a warning label for agent use. It says the model can loop through long reasoning without progress, may delete or overwrite files without confirmation, may misreport what it did, and does not fully stop reasoning at the lowest effort setting. The lab recommends capping output tokens and agent turns and running tools in a sandbox. Local runtime support for looped models is new, so speed depends on which engine you use, and the three passes cost more compute per token than the 3.4 billion active parameters suggest.

Our take: for a team choosing one open model for a 32GB Mac or a single workstation GPU, this and Qwen3.8-27B are now the pair to compare; Inkling 2 Mini edges ahead on coding agents, documents and screen reading, Qwen on factual recall, and only the Inkling comes from a lab whose larger model you cannot download.

Where it earns its place

  • A solo developer with a 32GB or larger Apple silicon MacBook

    Run a local coding agent without sending code to an API

    `ollama run inkling-2-mini` pulls the 20GB build, which Ollama lists for Claude Code, OpenCode and Hermes Agent. Cap agent turns and keep the working copy under version control, as the lab advises.

  • An operations team at an insurer processing scanned claims

    Extract fields from PDFs and screenshots on in-house hardware

    Its 92.8% on OmniDocBench v1.6 leads the lab's table. Served with vLLM or SGLang on one 32GB GPU, documents never leave the building and there is no per-page fee.

  • An ML team at a mid-sized software company

    Fine-tune a small vision model on internal tickets and own the weights

    Train LoRA adapters on Tinker at $0.75 per million tokens with 64K context, or with Unsloth, Miles or SkyRL on your own GPUs, then serve the result locally under Apache 2.0.

Before you commit

Setup
Download from Hugging Face or run `ollama run inkling-2-mini`; llama.cpp users can serve the ggml-org GGUF with `llama serve -hf ggml-org/Inkling-2-Mini-GGUF:IQ4_NL`. Recipes exist for vLLM, SGLang, TokenSpeed, MLX and Transformers.
Integrations
Transformers, vLLM, SGLang, TokenSpeed, NVIDIA Dynamo, llama.cpp and Llama App, Ollama, Exo, Unsloth; post-training with Miles and SkyRL; Tinker API for hosted fine-tuning and sampling.
Data and privacy
Apache 2.0 weights run entirely on your hardware, so nothing leaves it unless you use Tinker. Training data includes public web content, third-party data and synthetic data; training cutoff not published.
Cost at scale
Self-hosted, the cost is one 32GB GPU, DGX Spark or Mac per instance. On Tinker, 1 billion input and 100 million output tokens a month cost about $200 at list price.
Limits
1M-token context; images ideally 40 to 4,096 pixels a side; text output only. The IQ4_NL GGUF keeps the same top token as BF16 95.5% of the time, Q8_0 97.6%. Output quality drops outside English.

What’s good

  • Apache 2.0, with a 20GB official checkpoint and GGUF builds from 17.6GB
  • Ties Qwen3.8-27B at 20.1 on the AA Index v5.0 in the lab's table
  • 79.2% on SWE-bench Verified and 82.7% on MCP Atlas, per the model card
  • Text and image input with up to 1M tokens of context
  • Day-one support in llama.cpp, Ollama, vLLM, SGLang, MLX and Transformers

What’s not

  • All benchmark figures come from Thinking Machines
  • The lab warns of reasoning loops and unconfirmed destructive actions in agent use
  • 63.4% non-hallucination rate, below Qwen3.8-27B's 71.4%
  • Three passes per token make it heavier to run than 3.4B active implies

Best for

Developers and small teams who want a permissively licensed coding, document and screen-reading model that runs on one Mac or workstation GPU.

Inkling 2 Mini alternatives

Inkling 2 Mini: common questions

Is Inkling 2 Mini worth it?
Inkling 2 Mini is the open model Thinking Machines owed after holding back Inkling 2 Small: a 20GB download that runs on a laptop and keeps pace with Qwen3.8-27B. We rate it 4.3 out of 5.
How much does Inkling 2 Mini cost?
Free weights under Apache 2.0. Tinker: $0.15 per 1M input tokens ($0.015 cached), $0.50 per 1M output; training $0.75 per 1M tokens at 64K context, $1.50 at 256K.
What are the downsides of Inkling 2 Mini?
All benchmark figures come from Thinking Machines; the lab warns of reasoning loops and unconfirmed destructive actions in agent use; 63.4% non-hallucination rate, below Qwen3.8-27B's 71.4%; three passes per token make it heavier to run than 3.4B active implies.
Who is Inkling 2 Mini best for?
Developers and small teams who want a permissively licensed coding, document and screen-reading model that runs on one Mac or workstation GPU.
How much does Inkling 2 Mini cost for a team or in production?
Self-hosted, the cost is one 32GB GPU, DGX Spark or Mac per instance. On Tinker, 1 billion input and 100 million output tokens a month cost about $200 at list price.
Is Inkling 2 Mini safe for company data?
Apache 2.0 weights run entirely on your hardware, so nothing leaves it unless you use Tinker. Training data includes public web content, third-party data and synthetic data; training cutoff not published.
What are the best Inkling 2 Mini alternatives?
The closest alternatives we have reviewed are Qwen3.8-27B, Gemma 4, Inkling 2 Small, Inkling and Nemotron 3.5 Super VL.

Answers checked Nov 21, 2026.

Make Inkling 2 Mini? Show this rating on your siteInkling 2 Mini rated 4.3 out of 5 on ReviewRamble