Inkling 2 Mini is a 26.4-billion-parameter open-weight looped Mixture-of-Experts model released by Thinking Machines Lab on November 20, 2026. Licensed under Apache 2.0, its ~20GB mixed NVFP4 and MXFP8 checkpoint targets local agentic coding on Apple silicon with 32GB or more memory, 32GB+ VRAM GPUs, or NVIDIA DGX Spark.
What did Thinking Machines Lab announce on November 20, 2026?
Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, through an official announcement blog post, a same-day X launch thread, and open weights under an Apache 2.0 license on its Hugging Face model card. In the X post, Thinking Machines wrote: “Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.”
While the X post rounds the figure to “just 26B,” the official blog, model card, Inkling family page, and Tinker catalog all report 26.4B total parameters. Within the broader lineup of open-weight models in 2026, Inkling 2 Mini sits at the entry tier of a three-model family:
- Inkling 2 Mini: 26.4B total parameters, 3.4B active per pass, positioned as a laptop-scale looped transformer for agentic coding.
- Inkling 2 Small: 594B total parameters, 17B active, positioned as the flagship second-generation model for agentic coding and vision.
- Inkling: 975B total parameters, 41B active, covering text, images, and audio.
How does Inkling 2 Mini’s looped MoE architecture work?
Inkling 2 Mini combines a Mixture-of-Experts (MoE) design with weight-shared layer looping so a 26.4B model can spend extra compute per token without growing its memory footprint. On a single pass through its 24 decoder-only layers, each token routes to 8 of 128 experts across a hybrid of local and global attention layers, activating 3.4B parameters.
Those 24 layers are split into four six-layer blocks. The first and last blocks execute once, while the middle 12 layers loop three times with shared weights, producing 48 layer applications per token from 24 layers of weights. Each additional pass through the middle layers adds about 1.0B effective active parameters, bringing total weight activations to 5.4B per token. That explains why the family page lists 3.4B active parameters while the Tinker platform catalog lists 5.4B active parameters: one figure counts per-pass experts, and the other reflects all three looped passes.
Thinking Machines built Inkling 2 Mini by pruning and healing a larger sparse network. Training began with 8 of 384 experts active per layer before pruning two-thirds of the experts down to the 128 most-used experts per layer. The pruned model then underwent 1T tokens of “healing” via model distillation from a larger internal Inkling model, supervised fine-tuning, 2.2T tokens of reinforcement learning across 2,000 steps, and on-policy distillation for alignment.
Unlike post-training compressed checkpoints, Inkling 2 Mini was trained and deployed natively in mixed low precision: 20 layers in 4-bit NVFP4 and the final 4 layers in 8-bit MXFP8, avoiding post-training conversion loss as covered in our AI quantization explainer. The result is a single ~20GB checkpoint that accepts text and pixel images (40px to 4096px recommended) and outputs text in English plus general multilingual support, though the model card warns that Chinese testing showed clipped phrasing and mixed languages.
Official hardware requirements, context limits, and where to get Inkling 2 Mini
Inkling 2 Mini requires roughly 20GB for its native checkpoint and targets hardware with at least 32GB of unified memory or VRAM, according to official Thinking Machines Lab documentation as of November 21, 2026.
Official hardware guidance and the MacBook Pro demo
The official model card lists three supported local hardware targets:
- Apple silicon Macs: MacBooks, Mac minis, and Mac Studios with 32GB or more unified memory.
- Discrete GPUs and other platforms: 32GB or more VRAM recommended.
- Workstation hardware: 1x NVIDIA DGX Spark, alongside kernel acceleration on NVIDIA’s GB300 platform for NVFP4/MXFP8 execution.
Thinking Machines also published an offline “PlaneBench flight simulator demo” showing researchers running Inkling 2 Mini in airplane mode on a MacBook Pro (M5 Pro, 48GB) to build a playable flight simulator in a few minutes, with most footage sped up 15x. That clip illustrates one offline session rather than a minimum spec or a standardized speed benchmark.
Context limits and where to run the weights
Official documentation reports two distinct context limits depending on where you run the model. The Hugging Face model card states that Inkling 2 Mini supports a context window of up to 1M tokens, whereas the official Tinker model catalog currently caps hosted Inkling 2 Mini at 256K max context.
Developers can download the Apache 2.0 weights for Inkling-2-Mini and the research-only variable-loop variant TML-d24-varloop-research directly from Hugging Face. Thinking Machines lists open-source inference support across SGLang, vLLM, Ollama, Unsloth, NVIDIA Dynamo, TokenSpeed, and Llama App, while the model card also notes llama.cpp for quantized GGUF builds and Exo Labs on Apple silicon, plus SkyRL, Miles, and NVIDIA AutoModel for RL post-training. Hosted inference, fine-tuning, and interactive chat are available via Tinker and Tinker Playground. As of November 21, 2026, official public pages reviewed for this report list Tinker checkpoint storage at $0.10 per GB per month, note that training and inference rates vary by model and context length, and do not publish a complete per-token rate table for Inkling 2 Mini.
How does Inkling 2 Mini score on official benchmarks?
Thinking Machines Lab reports that Inkling 2 Mini leads open-weight models under 100B total parameters on DeepSWE v1.1 and Vals Index v2.2, while competing closely with Qwen3.8-27B and Nemotron 3.5 Super VL (124.5B total) across coding, browsing, and tool-use suites. All figures below come from Thinking Machines’ official blog and model card as of November 21, 2026, where evaluation harnesses vary by model.
| Benchmark (official harness note) | Inkling 2 Mini (5.4/26.4B) | Qwen3.8-27B (27.0/27.0B) | Nemotron 3.5 Super VL (12.0/124.5B) |
|---|---|---|---|
| AA Index (v5.0) | 20.1 | 20.1 | 13.6 |
| Vals Index (v2.2) | 46.5% | — | 29.8% |
| DeepSWE v1.1 (mini-swe-agent v2.4.6) | 48.2% | 42.2% | 24.7% |
| SWEBench Pro v2 Hard (mini-swe-agent v2.4.6) | 77.9% | 73.8% | 44.1% |
| SWEBench Verified (bash only) | 79.2% | 77.8% | 71.1% |
| Terminal Bench v2.1 (pi v0.83.0 for Mini) | 71.6% | 73.0% | 60.7% |
| PMPP-Hard (AUC, 0–10h, pi v0.83.0) | 455.82 | 262.50 | — |
| BrowseComp (Mini with compaction) | 79.0% | 61.6% | 78.9% |
| DeepSearchQA (Mini with compaction) | 85.3% | 79.7% | 77.7% |
| MCP Atlas Public | 82.7% | 77.3% | 67.7% |
| LongBench v2 | 66.9% | 52.3% | 62.1% |
| HLE-Diamond | 15.2% | 15.3% | 8.8% |
| CritPt v2 Full / Hard | 31.4% / 17.0% | 10.6% / 4.0% | 36.9% / 13.5% |
On early academic testing cited in the official blog, researchers at Stanford’s Hazy Research Lab and Scaling Intelligence Lab also measured Inkling 2 Mini at 0.703 on their “Intelligence per Watt” metric using an NVIDIA DGX Spark (128GB) in MXFP8 with low reasoning effort, ahead of Qwen3.6-35B-A3B (0.430), Gemma-4-26B-A4B (0.159), and Qwen3.5-27B (0.036) tested in NVFP4 and INT4.
What are Inkling 2 Mini’s limitations and monitorability findings?
Inkling 2 Mini trades broad factual recall and strict multi-turn instruction adherence for stronger agentic coding and tool execution, according to Thinking Machines Lab’s official disclosures.
Practical limitations and agent safeguards
The official blog states plainly: “Inkling 2 Mini is not the strongest small model for multi-turn chat or for instructions with many constraints: it scores 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, behind Qwen3.8-27B on both. It also trails larger models on reasoning-intensive benchmarks such as CritPt v2 and HLE-Diamond.” (For reference, the model card table separately lists Qwen3.8-27B at 30.1% on ComplexConstraints v1.1 and Muse Glimmer at 35.6%.)
The model card highlights additional constraints worth knowing before deploying local agents:
- Knowledge gaps: Looping cannot store extra facts in weights, leaving Inkling 2 Mini behind Inkling 2 Small on factual recall (AA-Omniscience v1.1 score -8.6, 63.4% non-hallucination rate) and more prone to confident inaccuracies without search tools.
- Debugging loops: Internal testing showed the model sometimes got stuck re-checking code or looping endlessly, prompting Thinking Machines to recommend strict output-token and agent-turn caps.
- Destructive tool risk: The model card explicitly warns that “It may take destructive actions, such as deleting or overwriting files, without confirmation.”
- Compute floor: Setting reasoning effort to minimum does not fully disable reasoning, and three middle-layer passes mean per-token compute exceeds what 3.4B active parameters imply, while local runtime support for looped models remains new.
Looping and chain-of-thought monitorability
Because looped transformers reason internally across passes without writing everything down, Thinking Machines released TML-d24-varloop-research (trained across 1 to 5 passes, primarily 3) to test monitorability. Without chain of thought, time-horizon capability rose from 16 seconds at one pass to 37 seconds at three passes, and serial N-Hop Lookup performance kept climbing up to 10 loops. On GPQA Diamond traces solved correctly at all five pass counts, average reasoning length dropped from 3,703 tokens at 1 pass to 1,661 tokens at 5 passes with unchanged sentence length and word choice, and agentic sabotage tests showed no upward trend as loops increased.
Thinking Machines summarized the result on X as “no meaningful drop in monitorability.” At the same time, both Inkling 2 Mini and the three-pass variable-loop model showed poor control over their chain of thought, following CoT instructions at most 0.1% of the time in 750–1,250-token traces. These observations describe behavior at 26.4B scale and do not certify safety for unsupervised deployment.
Who is Inkling 2 Mini actually built for?
Inkling 2 Mini is built for developers running local coding, search, and MCP tool-use agents on 32GB+ Apple silicon Macs, 32GB+ VRAM GPUs, or DGX Spark systems who want Apache 2.0 weights they can inspect or fine-tune. From my reading of the release notes, the strongest fit is structured software engineering tasks where tool outputs, tests, and search ground the model’s responses and turn limits keep loops bounded.
If you are deciding how to choose an AI model primarily for multi-turn conversational chat, dense factual Q&A without retrieval, or unmonitored file-system autonomy, larger models such as Inkling 2 Small or dedicated chat-focused alternatives remain safer matches.
Frequently asked questions about Inkling 2 Mini
Why does Thinking Machines list both 3.4B and 5.4B active parameters for Inkling 2 Mini?
Inkling 2 Mini activates 3.4 billion parameters on a single pass through its 24 layers, which matches the number on the Thinking Machines family page. Because its middle 12 layers loop three times per token, each extra pass adds roughly 1.0 billion effective active parameters. That brings total effective activations per token to 5.4 billion, which is the figure listed in Tinker.
Can you run Inkling 2 Mini on a laptop with 16GB of RAM?
Thinking Machines Lab does not list 16GB of RAM as a supported setup in its official documentation as of November 21, 2026. The official checkpoint is roughly 20GB in mixed NVFP4 and MXFP8 precisions, and official guidance specifies Apple silicon Macs with 32GB or more memory, GPUs with 32GB or more VRAM, or an NVIDIA DGX Spark.
What is the difference between Inkling 2 Mini’s 1M and 256K context limits?
The two numbers reflect local model capability versus current hosted limits. On Hugging Face, the official Inkling 2 Mini model card states that the architecture supports context windows up to 1 million tokens. On the official Tinker catalog page, hosted Inkling 2 Mini currently carries a 256K maximum context limit, matching other models on that platform.
Is Inkling 2 Mini truly open weight and free for commercial use?
Inkling 2 Mini is released under the permissive Apache 2.0 license on its official Hugging Face model card, allowing commercial and local use under Apache 2.0 terms. For hosted access on Tinker, official pages reviewed on November 21, 2026 list $0.10 per GB per month for checkpoint storage and do not publish a complete public per-token rate table.
How does Inkling 2 Mini compare to Inkling 2 Small?
Inkling 2 Small is a much larger model with 594 billion total parameters and 17 billion active parameters, while Inkling 2 Mini has 26.4 billion total parameters and 3.4 billion active per pass. Thinking Machines reports that Mini mirrors Small’s agentic coding strengths while trailing on factual recall, broad world knowledge, and heavy reasoning benchmarks.



