Inkling 2 Mini: A Free Open-Weight Coding Model That Runs on 20GB of Memory
Thinking Machines released Inkling 2 Mini on 20-21 November 2026. This free, open-weight model can run on a laptop with about 20GB of memory and leads open-weight coding tests.
As of 21 November 2026, the Thinking Machines announcement page and the Hugging Face launch post both list Inkling 2 Mini as open weights. The Hugging Face post also shows the model building a playable flight simulator with no internet connection, on a MacBook Pro with 48GB of memory. That is a laptop you can buy in a shop, but most students own an 8GB or 16GB machine, and the weights of this model are bigger than that.
The real question for a student is different. Is this a nice demo, or a coding model you can use for real work?
Inkling 2 Mini is a free, open-weight coding model from Thinking Machines. Its weights are about 20GB, so it needs a 32GB laptop, not a typical student machine. It leads open-weight models under 100B total parameters on agentic coding tests, according to the Hugging Face launch post, but it is weaker at chat and strict instructions. On an 8GB or 16GB laptop, use the hosted API instead.
What Thinking Machines released
Thinking Machines Lab announced Inkling 2 Mini late on 20 November 2026. It is an open-weight Mixture-of-Experts model with 26.4B total and 3.4B active parameters, per its announcement. "Total" is every weight stored on disk, while "active" is the small group of expert parts that does the work for one token.
Apache-2.0 is the license listed on the announcement page. It is a permissive license: you can use the weights in a college project or a paid product, as long as you keep the license notice. Read the license file on the model page before you ship anything, because that file is the official rule.
Compare these open weights with Alibaba's Qwen4 Max, a flagship you can only reach through an API. With Inkling 2 Mini, you download the file and it stays yours.
Looping makes 24 layers act like 48
The middle 12 of the model's 24 layers run three times for each token, so the model gets the depth of 48 layers without storing 48 layers of weights, as the Hugging Face post explains. A layer is one processing step the text passes through on its way to an answer.
Think of reading a hard page in a textbook. You read the easy parts once and re-read the tricky middle three times. You did not buy three copies of the book. You spent more time on the hard part, and looping does the same: more computing, same weights.
| Part of the model | Layers stored | Runs per token | Layers of work |
|---|---|---|---|
| Looped middle layers | 12 | 3 | 36 |
| All other layers | 12 | 1 | 12 |
| Total | 24 | - | 48 |
| Source: Hugging Face launch post | |||
This helps most with agentic coding. In agentic coding, the model plans, runs tools, reads errors and tries again by itself. Looping lets it think harder inside each step. It does not need to write out a long chain of thought, so the reasoning it does write stays short and readable.
There is an honest limit. Looping adds thinking, not facts. What a model knows still depends on its total size, so a small model remembers less than a big one. Inkling 2 Mini makes up for this with tool use and search: it looks things up instead of remembering them (Hugging Face).
Will it fit on your laptop? The weights are 20GB, so plan for 32GB
The weights take about 20GB, so a laptop with 32GB of memory is a realistic target, while 16GB machines will likely struggle (Hugging Face). The 32GB advice is my rule of thumb, not a tested result. By that rule, an 8GB or 16GB laptop cannot hold the 20GB of weights, a 24GB laptop is tight, and 32GB is the realistic target. Your operating system, browser and code editor need memory too while the model runs.
Two compact number formats store the weights: NVFP4 for 20 layers and MXFP8 for 4 layers, per the launch post. Shrinking a model after training, called quantization, usually costs some accuracy. This model was trained in these same formats, so it should not need a conversion step, which avoids the accuracy loss that usually comes with one.
That M5 Pro MacBook Pro from the opening is the only machine the demo shows. That is the tested setup. Anything smaller is your own experiment, and a useful one to report back on.
It also uses less power per answer than a similar model (Hugging Face).
Where Inkling 2 Mini beats other open models, and where it trails
It leads every open-weight model under 100B parameters on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%), but it trails on multi-turn chat and constraint-heavy prompts, according to the Hugging Face comparison. The rivals in that comparison are Qwen3.8-27B, Gemma 4 31B, Muse Glimmer and Nemotron 3.5 Super VL. All of them are small open-weight models.
| Area | Result | Verdict |
|---|---|---|
| Agentic coding | Top open model under 100B | Strong |
| Search and tool tasks (BrowseComp, DeepSearchQA) | Strong results | Strong |
| Whole Hugging Face comparison | Leads in 32 of 48 benchmarks | Strong |
| Multi-turn chat (MultiChallenge) | 50.0% | Weak |
| Many strict rules in one prompt (ComplexConstraints) | 33.4% | Weak |
| Hard reasoning (HLE-Diamond) | Behind some rivals | Weak |
| Source: Hugging Face launch post | ||
On the broader AA Intelligence Index v5.0, it scores 20.1 (Hugging Face). None of my sources compare it with closed flagship models, so I make no claim there.
Pros
- Coding agents that edit files, run tests and fix their own errors
- Research bots that search the web and pull results together
- Offline work on code you cannot send to a cloud service
Cons
- An everyday chatbot with long back-and-forth conversations
- Tasks that need an exact output format, like strict JSON or a word limit
- Puzzle-style reasoning questions
What it costs: $0 self-hosted or $0.075 per million tokens
Self-hosting the open weights costs nothing beyond your own hardware and electricity, and the hosted API on OpenRouter charges $0.075 per million input tokens and $0.30 per million output tokens. A token is a small piece of a word. Output costs four times as much as input, so long generated answers drive the bill. As a worked example, 50 coding prompts of about 2,000 tokens in and 1,000 tokens out use 100,000 input and 50,000 output tokens, which is about $0.0075 plus $0.015, roughly $0.02 in total (my own arithmetic from the listed prices).
The OpenRouter listing adds three details:
- Cache reads cost $0.015 per million tokens. A cache read is when you send the same text again, like your project files on every request.
- The hosted context window is 1M tokens, enough to hold a mid-sized codebase in one request, but a local run on a laptop will have a much smaller practical limit because memory runs out first.
- Prime Intellect and io.net serve the model.
Running it locally gives you two things the API cannot: no data leaves your machine, and there is no bill however many times you retry. Use this short test to choose:
- Go local if your laptop has the memory and your code is private, like internship work under an agreement.
- Go hosted if your laptop is too small or you need the full context window at server speed.
Run Inkling 2 Mini this weekend
Four day-0 routes work today: Hugging Face transformers, llama.cpp or Llama App, MLX on Apple silicon, and SGLang, per the launch post. Day-0 means those tools supported the model on release day.
Pick your route
Mac: MLX. Any laptop: llama.cpp or Llama App. Python users: transformers. A rented GPU server: SGLang.
Check your memory
Compare your RAM with the rule of thumb from the memory section. Close heavy apps first.
Download the weights
Follow the model link in the Hugging Face post and copy the exact model ID from there.
Run a first prompt
Start with one small, clear coding task and read the whole answer.
Here is the shape of a transformers script. I have not run this on my own laptop, so I show no output or speed here. I left the model ID as a placeholder on purpose: take it from the Hugging Face post, along with any extra loading settings it lists. A guessed ID just fails to download.
# pip install transformers accelerate
from transformers import pipeline
# Paste the exact model ID from the Hugging Face launch post.
model_id = "PASTE_MODEL_ID_FROM_HF_POST"
coder = pipeline("text-generation", model=model_id, device_map="auto")
messages = [{"role": "user", "content": "Build a small snake game in one HTML file."}]
result = coder(messages)
print(result[0]["generated_text"][-1]["content"])
A good first project works in Bengaluru and in Boston alike: ask for a small browser game or a to-do app. Open it, find what breaks, and paste the error back. Watching the model fix its own bugs shows you how it behaves as an agent.
What to do next with Inkling 2 Mini
Pick one small coding project, run it on Inkling 2 Mini, and note where it fails so you learn its limits firsthand. Keep it away from two jobs: your main chat assistant, and any task that needs an exact output format. Use a different tool for those.
If you also want flagship numbers for comparison, read my price breakdowns of Gemini 4 Hydrogen and GPT-6.2 Sol. For the bigger picture of where these tools leave your career, see how an early-career engineer stays ahead of AI.
Choose a small coding project this weekend, such as a browser game or a to-do app, run it on Inkling 2 Mini with one of the four setup routes, and write down where it succeeds and where it fails.