Lounge
Connecting…
Connecting…
View all

Inkling 2 Mini vs Inkling 2 Small: Specs, Local Use, Limits

Inkling 2 Mini explained: 26.4B open-weights looped MoE from Thinking Machines, how it differs from Inkling 2 Small, 32GB Mac setup, license and known limits.

AILatest6 min read2 viewsPublished Updated
Contents7 sections
  1. What is Inkling 2 Mini? The key specs
  2. Inkling 2 Mini vs Inkling 2 Small: what is different
  3. How to run Inkling 2 Mini locally or by API
  4. What to check before you deploy it
  5. Bottom line and checklist
  6. Frequently asked questions
  7. Sources and references

Key takeaways

  • Inkling 2 Mini, released November 20, 2026, is a Mixture-of-Experts model with 26.4B total and 3.4B active parameters (5.4B effective with looping), under Apache 2.0.
  • It is a looped transformer: the middle 12 of 24 layers run three times per token, which improves reasoning but not stored factual knowledge.
  • The ~20GB checkpoint (NVFP4 plus MXFP8) is supported on Apple silicon with 32GB+ memory, NVIDIA DGX Spark, or GPUs with 32GB+ VRAM.
  • Inkling 2 Small (594B total, 17B active) scores higher on most evaluations, including DeepSWE v1.1 (72.1% vs 48.2%), but is a server-class model.
  • The model card warns about destructive tool actions, reasoning loops, weaker non-English output, and multi-turn chat, so sandbox agent runs.

Inkling 2 Mini is a small open-weights model that Thinking Machines Lab released on November 20, 2026. It is a Mixture-of-Experts model with 26.4B total parameters and 3.4B active per token, packaged as a roughly 20GB checkpoint that runs on a Mac with 32GB of memory or an NVIDIA DGX Spark. It ships under Apache 2.0. Compared with the company's larger Inkling 2 Small (594B total parameters), Mini is far lighter and faster, but weaker on factual knowledge, complex instructions, and long multi-turn chat. This guide is based on the official announcement and the model card, as of November 21, 2026.

What is Inkling 2 Mini? The key specs

The defining feature is the looped transformer design. The model has 24 layers split into four blocks of six. The first and last blocks run once, while the two middle blocks (12 layers) run three times for every token. That gives 48 layer applications per token from 24 stored layers. Each extra pass adds about 1.0B effective active parameters, which is how the model card arrives at 5.4B effective active parameters.

ItemInkling 2 Mini (published)
Release dateNovember 20, 2026
Architecture24-layer looped MoE, 8 of 128 experts per layer
Parameters26.4B total, 3.4B active, 5.4B effective active
Precision and sizeNVFP4 (20 layers) + MXFP8 (last 4), ~20GB
ModalitiesText and image in, text out
Context windowUp to 1M tokens
LicenseApache 2.0, weights on Hugging Face

Thinking Machines says looping lets the model reason more while writing fewer tokens, which helps on agentic coding. It reports 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2, and claims the top spot among open-weights models under 100B total parameters on those benchmarks. On the Artificial Analysis Index v5.0 it reports 20.1, tied with Qwen3.8-27B. These are vendor-reported numbers, and the announcement notes that evaluation setups vary by model.

Looping improves reasoning, not memory. Thinking Machines states that knowledge is mostly a function of total parameter count, so Mini relies on search and tool use to close the knowledge gap.

Alongside Mini, the company released TML-d24-varloop-research, a research variant that can run one to five loop passes. It exists mainly so researchers can study how looping affects chain-of-thought monitorability, not as a production model.

Inkling 2 Mini vs Inkling 2 Small: what is different

Inkling 2 Mini vs Inkling 2 Small: what is different

Both are second-generation Inkling models, but they target different deployments. Inkling 2 Small was announced on November 5, 2026 with closed weights at first; a November 12 update made the weights downloadable on Hugging Face. Mini launched with open weights and Tinker access on day one.

Announced

Inkling 2 Mini
Nov 20, 2026
Inkling 2 Small
Nov 5, 2026 (weights Nov 12)

Total parameters

Inkling 2 Mini
26.4B
Inkling 2 Small
594B

Active parameters

Inkling 2 Mini
3.4B (5.4B effective)
Inkling 2 Small
17B

Design

Inkling 2 Mini
Middle 12 layers looped 3x
Inkling 2 Small
Standard MoE, variable effort

AA Index

Inkling 2 Mini
20.1 (v5.0)
Inkling 2 Small
31

DeepSWE v1.1

Inkling 2 Mini
48.2%
Inkling 2 Small
72.1%

Typical deployment

Inkling 2 Mini
32GB Mac, DGX Spark, local
Inkling 2 Small
API providers, large GPU servers

The model card is direct about the gap: Mini is less capable than Small across most evaluations, with the largest differences on factual recall and knowledge-heavy questions. One exception is kernel optimization. On PMPP-Hard, Mini's reported AUC of 455.82 comes close to the 494.27 listed for Small in its own announcement.

When Mini makes sense

  • Offline or on-device coding agents where data cannot leave the machine
  • Running many persistent agents on one box; the announcement demos 20 villager agents on a single offline MacBook
  • Speed-sensitive inference where you can add search or retrieval for facts

When Small is the better fit

  • Closed-book question answering that depends on stored knowledge
  • Finance, spreadsheet, and document analysis, which Small's announcement highlights
  • Instructions with many constraints and long multi-turn conversations

How to run Inkling 2 Mini locally or by API

  1. 1

    Try it in the browser first

    The Tinker Playground offers inference-only chat access, so you can test real prompts before downloading anything.

  2. 2

    Check your hardware

    Supported setups are one NVIDIA DGX Spark, Apple silicon Macs with 32GB or more (MacBook, Mac mini, Mac Studio), or other GPUs with 32GB or more VRAM recommended. Budget for context memory on top of the 20GB weights, especially with long prompts.

  3. 3

    Pick a runtime

    On Apple silicon, use Ollama or Exo Labs. On NVIDIA, the model card lists SGLang, vLLM, Unsloth, NVIDIA Dynamo, and TokenSpeed. Quantized GGUF builds are available through Unsloth and llama.cpp.

  4. 4

    Use the API or fine-tune

    Tinker supports inference and fine-tuning, and third-party inference providers also host the model. Check current per-token rates on the Tinker Models & Pricing page before you scale.

Expect some rough edges. The model card says local runtime support for looped models is new and speed varies by runtime and hardware. Because the middle layers run three times, compute per token is higher than the 3.4B active count suggests. The speed chart in the announcement is a theoretical ceiling assuming 614 GB/s bandwidth, BF16 weights, and batch size one, so real throughput on your laptop will differ.

What to check before you deploy it

License and terms

Apache 2.0 covers the weights, but the model card also links a Model Acceptable Use Policy. Hosted access adds the provider's own terms and data-handling rules, which matters if you plan to send company code or customer data.

Known limitations from the model card

  • Weaker on multi-turn chat and multi-constraint instructions: 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, both behind Qwen3.8-27B.
  • More likely than larger models to state incorrect information with confidence.
  • Can get stuck in long reasoning loops while debugging, so set output-token and agent-turn limits.
  • May delete or overwrite files without confirmation and may misreport the outcome of its work.
  • Output quality can drop outside English, and minimum reasoning effort does not fully disable reasoning.

Run agent workloads in a sandbox and require human confirmation before destructive or irreversible actions. Thinking Machines also advises against medical, legal, or safety-critical use without extra fine-tuning, domain validation, and human oversight.

Launch-week benchmark claims tend to be followed by independent user tests. To see how developers are actually reacting, you can follow the discussion flow on blog trends.

Bottom line and checklist

  1. Confirm 32GB or more of unified memory or VRAM.
  2. Decide whether your workload is reasoning-heavy or knowledge-heavy; add retrieval or consider Small for the latter.
  3. Test the languages you actually serve.
  4. Wrap agents in a sandbox with output limits and human approval.
  5. Read the provider price list and terms before production use.

On the published numbers, Inkling 2 Mini is one of the strongest open-weights coding and agent models you can run on a single laptop. Those numbers are self-reported by Thinking Machines with setups that differ across models, so reproduce them on your own tasks before committing to it.

Frequently asked questions

Is Inkling 2 Mini open source and free for commercial use?
The weights are published on Hugging Face under the Apache 2.0 license, which generally permits commercial use, modification, and redistribution. Thinking Machines also links a separate Model Acceptable Use Policy, and if you call the model through Tinker or a third-party inference provider, that service's terms and per-token pricing apply on top.
Can Inkling 2 Mini run on a 16GB laptop?
Not in its official configuration. The model card lists Apple silicon with 32GB or more and GPUs with 32GB or more VRAM. The native checkpoint is about 20GB before context memory. Quantized GGUF builds exist through Unsloth and llama.cpp, but going below the native precision gives up the no-conversion-loss advantage and may reduce quality.
Should I pick Inkling 2 Mini or Inkling 2 Small?
Choose Mini for fast local inference, offline coding agents, or many persistent agents on one machine. Choose Small when you need stronger factual recall, finance and document work, or complex multi-constraint instructions. Small is offered by API providers such as Together, Fireworks, Baseten, Nebius, Modal, and Cloudflare, which is the practical route for most users.
What is a looped transformer?
It is a transformer that reuses some layers several times for each token instead of passing through every layer once. Inkling 2 Mini stores 24 layers but performs 48 layer applications per token. This adds serial reasoning depth without adding parameters, which is also why Thinking Machines studied whether it makes chain-of-thought monitoring harder.

Sources and references

  1. 1Introducing Inkling 2 Mini - Thinking Machines Lab (Nov 20, 2026)(opens in a new tab)thinkingmachines.ai
  2. 2Inkling 2 Mini Model Card - Thinking Machines Lab(opens in a new tab)thinkingmachines.ai
  3. 3Introducing Inkling 2 Small - Thinking Machines Lab (Nov 5, 2026)(opens in a new tab)thinkingmachines.ai
  4. 4Models & Pricing - Tinker Documentation(opens in a new tab)tinker-docs.thinkingmachines.ai