Field guide Inkling 2 Mini

Checked

Unofficial guide

Inkling 2 Mini Ollama

Last updated

Ollama publishes Inkling 2 Mini as inkling-2-mini. The library page's entire get-started block is one command, and the model card lists Ollama as a supported runtime for Apple silicon.

This is an unofficial guide. The commands in the blocks below are copied from Ollama's library page, not written for this site.

Run it

From ollama.com/library/inkling-2-mini
ollama run inkling-2-mini

Ollama's readme on that page says the model has 26 billion parameters with 5.4 billion active per token, loops 12 of its 24 layers three times, and is small enough for a Mac with 32 GB or more of memory. The 5.4 billion figure is the model card's effective count after three passes, not a different model. The card's base active count is 3.4 billion. Both are explained on the overview.

The library page does not name a minimum Ollama version. The model card says support for looped models in local runtimes is new, so use a current Ollama rather than a months-old install. There is no version number on the library page to quote.

Tags and what they weigh

The tags page lists three names. inkling-2-mini:latest and inkling-2-mini:26b-a5b-nvfp4 are both shown at 20 GB, with a 1 million token context and text plus image input. The NVFP4 tag is marked MLX, and on the day this guide was checked it shared the digest d183b1321cb5 with latest. A third tag, inkling-2-mini:26b-a5b, is listed beside them. Its digest was not clearly printed in that tags view, so this guide does not claim it is byte-identical.

As listed on Ollama's tags page. All three are the library model, not Unsloth GGUF quants.
Tag Size shown Context Note
inkling-2-mini:latest 20 GB 1M The name ollama run uses if you omit a tag
inkling-2-mini:26b-a5b Listed with the 20 GB builds 1M "a5b" matches the 5.4B effective figure, rounded
inkling-2-mini:26b-a5b-nvfp4 20 GB 1M Marked MLX. Native 4-bit, same story as the announcement

There is no smaller Ollama tag on that page. If 20 GB does not fit, use a GGUF quant from the GGUF page instead of hunting for a hidden Ollama tag. Unsloth documents unsloth run and llama-server for those files, not an ollama run hf.co/... line.

Call it from a script

Ollama prints these three clients on the library page. The HTTP example talks to port 11434 on localhost.

HTTP
curl http://localhost:11434/api/chat \
  -d '{
    "model": "inkling-2-mini",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'
Python
from ollama import chat

response = chat(
    model='inkling-2-mini',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)
JavaScript
import ollama from 'ollama'

const response = await ollama.chat({
  model: 'inkling-2-mini',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Coding agents

The same library page has launch commands for four agent apps. They are the documented way to point those tools at this model. They assume the apps are installed and that Ollama is already serving.

Ollama launch commands
ollama launch claude --model inkling-2-mini
ollama launch opencode --model inkling-2-mini
ollama launch hermes --model inkling-2-mini
ollama launch openclaw --model inkling-2-mini

The model card's agent warning still applies: Mini may delete or overwrite files without confirmation, and it may not report the outcome of its work accurately. Cap agent turns. The card says internal testing saw the model get stuck re-checking itself while debugging code.

Sampling

The NVFP4 tag's details on Ollama include a generation config with "do_sample": true and "temperature": 1.0. Unsloth, for the GGUF path rather than this library build, publishes temperature 1.0, top-p 1.0, top-k 0, and min-p 0.0. This guide does not add extra Ollama flags on top of the library page. If you need those samplers on a quant, use the llama.cpp command Unsloth prints.

What this build is for

Ollama's feature list says the build runs locally, is trained and shipped in NVFP4 so there is no extra quantization loss, and is Apache 2.0. Benchmarks they repeat from the lab: 79.2% on SWE-bench Verified, 48.2% on DeepSWE, 82.7% on MCP Atlas, 85.3% on DeepSearchQA, 76.5% on ScreenSpot Pro, and 92.8% on OmniDocBench. Those are the same model-card numbers. Harness notes live on the overview, because a bare percentage hides whether the run was bash-only or used a specific agent.

Context on the tag is 1 million tokens. That is the model's maximum, matching model_max_length 1048576. It is not a promise that a 32 GB Mac can fill the whole window. Unsloth's own llama.cpp example uses 32768 for that reason. If Ollama starts swapping or gets killed, shorten the context in Ollama's settings before you blame the weights. The troubleshooting page covers out-of-memory kills.

Mac, specifically

For Apple Silicon install order, MLX as an alternative, and what an Intel Mac is not promised, use Run Inkling 2 Mini on Mac. The short version: 32 GB unified memory is the floor Thinking Machines and Ollama both print, and the announcement's offline demo used a 48 GB M5 Pro.

Common questions

What is the Inkling 2 Mini Ollama command?

The Ollama library page's get-started command is `ollama run inkling-2-mini`. The same page shows a chat API on http://localhost:11434/api/chat.

How much memory does the Ollama build need?

Ollama's readme says the model runs on a Mac with 32 GB or more of memory, or a DGX Spark. The tags page lists the builds at 20 GB with a 1 million token context. Thinking Machines' model card gives the same 32 GB Apple silicon floor for the open checkpoint.

Does Ollama's Inkling 2 Mini accept images?

The library page marks the model as text and image input, and Ollama's search listing tags it vision, tools, and thinking. The page does not print a separate image-attach command, so this guide does not invent one.

Sources

  1. Ollama library page for inkling-2-mini
  2. Ollama tags for inkling-2-mini
  3. Inkling 2 Mini model card Hardware, license, architecture, evaluations, and limitations.
  4. Thinking Machines Lab, “Introducing Inkling 2 Mini” Announcement dated 20 November 2026.
  5. Unsloth, Inkling 2 Mini run guide GGUF sizes, memory ranges, sampler settings, and llama.cpp commands.