- Latest News about Uncensored AI
- What Is Inkling 2 Mini? Thinking Machines’ 26.4B Looped Model You Can Run on a Laptop (2026)
What Is Inkling 2 Mini? Thinking Machines’ 26.4B Looped Model You Can Run on a Laptop (2026)
Inkling 2 Mini is Thinking Machines Lab’s 26.4-billion-parameter open-weights model, released on November 20, 2026, for text-and-image input and text output. Its distinguishing feature is a looped transformer: it reuses its middle layers three times per token. The official mixed-precision checkpoint is about 20GB, and the model card supports Apple silicon devices with at least 32GB of memory—making local laptop inference a documented option, not a promise that every laptop can run it well.
The attraction is practical: a model intended for coding, tool use, and general multimodal work without requiring the hardware footprint of Inkling 2 Small. The trade-offs include runtime maturity, extra computation from looping, and explicitly documented reliability limitations. This introduction is based on official sources checked on November 21, 2026; we have not independently tested its speed, memory consumption, or benchmark performance.
What is Inkling 2 Mini?
Thinking Machines describes Mini as a general-purpose multimodal model for applications such as coding assistants, chatbots, retrieval-augmented generation, and agentic systems. Its official model card confirms:
- Release: November 20, 2026.
- Weights license: Apache 2.0.
- Inputs and output: text and images in; text out.
- Architecture: a 24-layer decoder-only transformer with sparse mixture-of-experts feed-forward layers and repeated middle layers.
- Parameter count: 26.4B total and 3.4B active, with approximately 5.4B effective active parameters per token when repeated passes are counted.
- Context: up to one million tokens.
- Distribution: downloadable weights on Hugging Face, alongside access through Tinker.
“Mini” is a relative name. A roughly 20GB checkpoint still requires a capable device, compatible software, and memory beyond the bare weight storage. It is not a tiny model meant for any ordinary low-memory notebook.
How does the looped transformer work?
The same middle layers are reused
The card describes 24 layers of weights. The middle 12 layers run three times with shared weights, yielding 48 layer applications per token: the 12 layers outside the loop run once, while the middle 12 contribute 36 applications.
That is different from storing 48 independent layers. Reusing weights lets the model apply more internal computation without increasing the stored layer count in the same way.
Thinking Machines argues in its launch post that this can improve reasoning quality while using fewer generated reasoning tokens. It is an architectural approach, not proof that every task becomes faster or more accurate.
Why both 3.4B and 5.4B active figures appear
Inkling 2 Mini uses mixture-of-experts routing: each token is routed to eight of 128 experts in a layer. Only part of the model is active for a particular token, which is why the active count is smaller than the 26.4B total.
The card’s 3.4B active figure does not include the repeated computational work in the same way as its approximately 5.4B effective figure. Each extra pass through the looped layers adds about 1.0B effective active parameters; three passes add two such repeats.
These are not three different checkpoint sizes. In particular, do not treat Mini as a normal 3.4B model when estimating computation or as a 5.4B-total model when estimating storage.
Does looping mean configurable thinking depth?
Not automatically. The released Mini runs the middle layers three times per token. Thinking Machines also released a separate variable-loop research variant for experiments comparing different looping depths.
Results from that research variant should not be mistaken for an adjustable loop-count feature supported by every Mini deployment. The model card also warns that setting reasoning effort to its minimum does not fully disable reasoning.
Can you really run Inkling 2 Mini on a laptop?
Yes, on supported hardware—but the memory qualifications matter. The official card lists:
- Apple silicon: MacBooks, Mac minis, and Mac Studios with 32GB or more memory.
- NVIDIA DGX Spark: one device is listed as supported.
- Other GPUs and platforms: possible, with 32GB or more VRAM recommended.
The launch includes a demonstration on a 48GB M5 Pro MacBook Pro running offline. That is Thinking Machines’ demonstration, not our replication, and it does not establish equivalent performance on every 32GB Mac.
About 20GB is not a universal runtime-memory requirement
The model card calls the distributed checkpoint about 20GB in mixed NVFP4 and MXFP8 precision. The launch describes a roughly 20GB model memory footprint. Neither statement should be read as “20GB of installed system memory is sufficient for any workload.”
Actual deployment also needs room for runtime allocations, context/KV cache, working buffers, the operating system, and other applications. These vary with the runtime and workload. A different GGUF quantization is also a different artifact from the native mixed-precision checkpoint.
The sensible purchasing threshold is therefore the published hardware guidance plus headroom—not simply the download size. We have not measured a minimum-memory configuration ourselves.
One-million-token context is a capability ceiling
The card specifies support for up to 1M tokens. That does not show that a maximum-length conversation will fit comfortably on a baseline 32GB laptop or run at an acceptable interactive speed.
When testing locally, begin with a modest context limit and a representative task. Increase the limit while observing memory and latency. Keep model-level context support, runtime implementation, and device capacity separate.
Which software can run it?
The official card lists a broad set of local inference options:
- Hugging Face, SGLang, vLLM, TokenSpeed, and NVIDIA Dynamo.
- Unsloth, including quantized GGUF versions.
- llama.cpp for quantized GGUF versions.
- Llama App.
- Ollama and Exo Labs on Apple silicon.
It also lists Miles and SkyRL for post-training. The launch mentions additional ecosystem support, but a support announcement is not a guarantee that an older installed version understands this architecture or its weight format.
Before downloading, check the current runtime-specific instructions for Inkling 2 Mini, your operating system, and the exact checkpoint. A generic transformer command or recipe for Inkling 2 Small may not be appropriate. Because native NVFP4/MXFP8 weights and converted GGUF files differ, confirm which artifact your chosen runtime expects.
We are not publishing an untested universal installation command. A short, correct runtime-specific recipe is more useful than a plausible command that silently targets a different model or precision.
Hosted access: Tinker and the Playground
Thinking Machines says Mini is available on Tinker for inference and fine-tuning, and through the Tinker Playground for chat. The card directs developers to the Tinker Cookbook and the tml-renderers package for integration.
This gives two distinct routes:
- Local weights: manage the device, runtime, model files, and application behavior yourself.
- Hosted Tinker access: use the provider’s service rather than fitting the model into your own memory budget.
Do not assume that the weight license makes hosted inference free, that every account has identical service limits, or that a local offline demonstration describes the hosted service’s data handling. Check current service pricing and terms separately; this article does not establish a verified Tinker price or quota.
Inkling 2 Mini versus Inkling 2 Small
The names can be misleading: Small is much larger than Mini. The official Small model card confirms a 594B-total, 17B-active MoE model.
Mini: a local-compute option
Mini has 26.4B total parameters, looped middle layers, and an approximately 20GB mixed-precision checkpoint. The official hardware guidance includes 32GB-and-up Apple silicon devices.
Its purpose is to bring useful general and agentic capability into a more accessible deployment envelope—not to guarantee the same factual depth or reliability as Small.
Small: a different hardware class
Small’s card describes a main NVFP4 checkpoint requiring a GPU cluster with at least 500GB of aggregated VRAM. Its alternative MXFP8 checkpoint requires at least 800GB.
Those are published deployment requirements, not a comparison inferred from parameter counts. They explain why a Mini laptop demonstration should not be presented as evidence that Small is laptop-sized too.
Capability is not interchangeable
Mini’s own card says it is less capable than Small across most evaluations, with the largest gaps in factual recall and knowledge-heavy questions. Both accept text and images and generate text, but shared modalities do not establish quality parity.
Choose based on your workload and available compute, not the assumption that “newer Mini” means “Small with no compromises.”
What do the benchmarks actually show?
Thinking Machines reports 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2, highlighting Mini’s performance among open-weights models below 100B total parameters. These are vendor-published results, not scores we reproduced.
The card includes evaluation harness and version information; for DeepSWE it names mini-swe-agent v2.4.6. That context matters because tool access, scaffolding, token budgets, and benchmark revisions can change outcomes.
There are also important caveats:
- The launch notes that evaluation setups vary by model.
- Its charted decoding speed is a theoretical ceiling at 614GB/s with BF16 weights and batch size one, including repeated reads across loops—not a measured laptop tokens-per-second promise.
- A coding score does not establish factual reliability, multilingual quality, or safe autonomous tool use.
- A successful vendor demo is one example, not a guarantee for your repository or device.
Use the results as a reason to evaluate Mini, not a substitute for running your own relevant tests.
Limitations and a safe first evaluation
The model card is unusually practical about failure modes. It warns about hallucination, weaker factual recall than Small, repetitive reasoning, and lower quality outside English. It also warns that tool-using agents may delete or overwrite files without confirmation and may inaccurately report their work.
A sensible initial evaluation is therefore deliberately constrained:
- Use a disposable copy of a project. Start with a small task and known tests, not your only working repository.
- Limit output tokens and agent turns. The card specifically recommends limits because reasoning can become repetitive.
- Require approval for destructive actions. Do not give a newly configured agent unrestricted access to production files, credentials, deployment controls, or payments.
- Measure the whole workflow. Record the runtime, artifact/precision, context length, memory use, time to first response, completion time, and whether the result passes checks.
- Verify claims independently. A model saying “tests passed” is not evidence unless the actual test output confirms it.
- Try your real input types. Evaluate representative documents or images, and your intended language, rather than relying only on English coding scores.
Open weights and Apache 2.0 licensing do not mean the model is unfiltered, risk-free, or accurate enough for unsupervised high-stakes decisions. Thinking Machines documents alignment work and recommends downstream safeguards. This is a product introduction, not an adult-content or jailbreak evaluation.
Who should consider it?
Mini is worth evaluating if you have supported local hardware, want control over deployment, and can validate a compact model against your own coding, document, or assistant tasks. It is less compelling if you need guaranteed frontier-level factual accuracy, a stable untested agent with broad permissions, or zero local setup effort.
If your immediate need is simply a browser-based chat workflow while you evaluate local options, you can explore HackAIGC Chat. Compare its current features and policies with your requirements. This is not a claim that HackAIGC hosts Inkling 2 Mini, runs offline, matches its capabilities, or shares its licensing and data-handling model.
FAQ
Is Inkling 2 Mini open source?
The official release provides open weights under Apache 2.0. That verifies the weights’ stated license; it does not by itself establish that the full training dataset and training pipeline are available.
Is it a 3.4B model or a 26.4B model?
It has 26.4B total parameters and 3.4B active. Counting the extra passes through its looped layers gives about 5.4B effective active parameters per token. Total storage and per-token computation are different quantities.
Will it run on a 16GB MacBook?
That is not supported by the hardware guidance reviewed here. The official card specifies Apple silicon with at least 32GB memory. Different quantizations need their own verified requirements; do not infer compatibility from a smaller file alone.
Can it generate images?
The card describes text-and-image input and text output. Multimodal understanding is not the same as native image generation.
Does it replace Inkling 2 Small?
Not without trade-offs. Mini targets a much smaller deployment footprint, while its card acknowledges lower capability than Small across most evaluations, especially factual and knowledge-heavy tasks.
Have you tested its local performance?
No. Hardware support, specifications, and scores here come from official documentation. Any deployment decision should include a runtime- and workload-specific test.
Sources and verification
Checked November 21, 2026:
- Thinking Machines Lab: Introducing Inkling 2 Mini, November 20 release, looping explanation, vendor results, demonstration, and ecosystem access.
- Official Inkling 2 Mini model card and weights, license, architecture, hardware/software support, modalities, context, and limitations.
- Official Inkling 2 Small model card, independently checked parameter counts and deployment requirements for comparison.
Runtime support and hosted availability can change. Confirm the current instructions for your chosen artifact before installing or purchasing hardware.
