Unofficial guide
Inkling 2 Mini
Last updated
Inkling 2 Mini is an open-weights multimodal model from Thinking Machines Lab. The announcement and the model card are both dated 20 November 2026. It is a looped mixture-of-experts built for local hardware, with its published strengths in long-horizon agentic coding, tool use, and visual work.
This page is an unofficial field guide. It repeats figures only when a primary source states them, and it links those sources at the bottom. It is not a Thinking Machines product.
- Released
- 20 Nov 2026
- Parameters
- 26.4B total, 3.4B active
- With 3 loop passes
- 5.4B effective active
- Checkpoint
- About 20 GB, NVFP4 and MXFP8
- Context
- 1,048,576 tokens
- License
- Apache 2.0
The parameter line is the model card's. The context length is model_max_length in the published config.json. The license on the model card is Apache 2.0, and the card also points at Thinking Machines' acceptable use policy.
Quick start
Pick the path that matches the machine in front of you. Each card goes to a longer page with the command and the source it came from.
ollama run inkling-2-mini. The published tags are about 20 GB. Apple Silicon and MLX The Hugging Face post shows MLX loading the original NVFP4 and MXFP8 checkpoint, vision included. GGUF and llama.cpp Unsloth's recommended quant is UD-Q4_K_XL at 17.63 GB. It needs a llama.cpp build that understands the loop. ollama run inkling-2-mini That one line is the whole official Ollama quick start. It is documented for a Mac with 32 GB or more of memory, or a DGX Spark. It is not the documented path for a consumer NVIDIA card. Those cards are covered under download and GGUF.
What it is
The model card calls Inkling 2 Mini a general-purpose multimodal model: text and images in, text out. Images do best when each side is between 40 and 4096 pixels. It is meant for agentic and tool-using apps, coding assistants, chat, and retrieval, in English and with general multilingual support. Output quality outside English can be lower. The card specifically mentions clipped phrasing and mixed languages in Chinese testing.
The announcement is more specific about where the model leads. Thinking Machines says it does best on long-horizon agentic work, and that among open-weights models under 100 billion total parameters it leads on DeepSWE v1.1 at 48.2% and Vals Index v2.2 at 46.5%. Their chart caption says evaluation setups vary by model. Treat the scores as the lab's reported numbers, not as an independent re-run.
Ollama's own readme describes the same model as built for coding, agentic work such as computer use, design that needs visual reasoning, and finance. That matches the model card's intended uses. It does not make the model a specialist that can only code.
Why people quote two active-parameter counts
Both numbers are in the model card, and they are not rivals.
- 26.4 billion is the total parameter count.
- 3.4 billion is the active count before the extra loop passes.
- Each extra pass through the looped layers adds about 1.0 billion effective active parameters.
- The model runs three passes, so the card's effective figure is 5.4 billion active parameters per token.
Ollama rounds the total to 26 billion and quotes 5.4 billion active. Unsloth says "26.4B MoE (5.4B active)". The Hugging Face blog rounds further, to about 26 billion total and 3 billion active, then explains the loop separately. When you compare models, say which of these you mean. The comparison with Inkling 2 Small uses the model-card pair, 26.4B and 3.4B, and notes the 5.4B effective figure beside it.
How the loop is wired
config.json sets num_hidden_layers to 24, hidden_size to 2304, and a loop_config with core_start 6, core_end 18, and num_passes 3. The core is layers 6 through 17. Those weights run three times. Layers 0–5 and 18–23 run once. That is 48 layer applications from 24 stored layers, which is also how the announcement describes its four blocks of six.
Layers 0–5
Prelude
Runs once. Its output feeds every pass.
Layers 6–17
Core ×3
Shared weights. Passes 2 and 3 also read the previous pass.
Layers 18–23
Coda
Runs once. The last four layers are the MXFP8 block.
Each token is routed to 8 of 128 experts. n_shared_experts is 0. Expert intermediate width is 1152. Attention has 24 query heads and 8 key-value heads. Twenty layers are local, with a sliding window of 1024 tokens. The four layers missing from local_layer_ids are 5, 11, 17, and 23, one full-attention layer in each block of six. The vision tower is 23 layers wide at hidden size 1024.
The announcement says most weights were trained and are deployed in NVFP4, with the last four layers in MXFP8, so there is no extra quantization step on the official checkpoint. The total memory footprint they quote is about 20 GB.
Scores the lab published
These are Thinking Machines' reported scores from the model card, not measurements made for this site. The closest model they table next to Mini on coding is Qwen3.8-27B (27B active and total in that row). Harnesses are not the same on every row. The full table also includes Gemma 4 31B, Muse Glimmer, Nemotron 3.5 Super VL, and several others.
| Benchmark | Setup noted on the card | Inkling 2 Mini | Qwen3.8-27B |
|---|---|---|---|
| DeepSWE v1.1 | mini-swe-agent v2.4.6 | 48.2% | 42.2% |
| SWE-bench Pro v2, Hard | mini-swe-agent v2.4.6 | 77.9% | 73.8% |
| SWE-bench Verified | bash only | 79.2% | 77.8% |
| Terminal-Bench v4.0 | mini-swe-agent v2.4.6 | 5.6% | 5.6% |
| Terminal-Bench v2.1 | Mini measured in pi v0.83.0 | 71.6% | 73.0% |
| BrowseComp | Mini with compaction | 79.0% | 61.6% |
| MCP Atlas | Public | 82.7% | 77.3% |
| AA Index v5.0 | Third-party index | 20.1 | 20.1 |
| Vals Index v2.2 | Third-party index | 46.5% | Not filled in |
| AA-Omniscience v1.1 | Score, then non-hallucination rate | -8.6, 63.4% | -6.4, 71.4% |
| ScreenSpot Pro | Original, then with Python | 76.5% / 87.7% | 57.6% / 78.9% |
| OmniDocBench v1.6 | Document reading | 92.8% | 91.5% |
On factuality, Mini does not pull away from other models its size. The announcement says knowledge is mostly a function of total parameter count, and that extra loop passes did not help knowledge-heavy benchmarks such as AA-Omniscience, SimpleQA Verified, SimpleVQA, and WorldVQA. Mini's AA-Omniscience score is -8.6, behind Qwen3.8-27B's -6.4 and Nemotron 3.5 Super VL's -2.9 in that table. The lab's point is that Mini is supposed to close the knowledge gap by searching and calling tools, which is why BrowseComp, DeepSearchQA, and MCP Atlas are the rows they highlight.
A separate efficiency note, with the lab's own caveat: researchers at Stanford's Hazy Research Lab and Scaling Intelligence Lab measured intelligence per joule and, the announcement says, scored Mini at 0.703 on an NVIDIA DGX Spark at MXFP8 with low reasoning effort. The comparison numbers in that paragraph are 0.430 for Qwen3.6-35B-A3B and 0.159 for Gemma-4-26B-A4B, both at NVFP4, and 0.036 for Qwen3.5-27B at INT4. The precisions are not the same. The paper they cite is "Intelligence per Watt", arXiv:2511.07885.
Limits worth knowing before you install it
- The model card says Mini is less capable than Inkling 2 Small across most of their evaluations, with the largest gaps on factual recall.
- In internal testing it sometimes got stuck re-checking its own work, especially while debugging. They recommend limits on output tokens and agent turns.
- With tools it is less reliable than larger models at following detailed instructions. It may delete or overwrite files without confirmation. Run agents in a sandbox.
- The announcement says it is not the strongest small model for multi-turn chat or for instructions with many constraints. They give MultiChallenge at 50.0% and describe that as behind Qwen3.8-27B. On ComplexConstraints v1.1 the limitations paragraph also says Mini is behind Qwen, while the score table lists Mini at 33.4% and Qwen3.8-27B at 30.1%. The table and the paragraph disagree. This guide does not pick a winner past printing both.
- Support for looped models in local runtimes is new. Speed depends on the runtime and the hardware. There is no official tokens-per-second table for a specific Mac in the sources checked for this site.
The flight-simulator demo in the announcement ran offline on a MacBook Pro (M5 Pro, 48 GB). That is a demonstration machine. The supported-hardware line is 32 GB or more of Apple silicon memory, not 48 GB.
Runtimes named by the lab
The model card splits local inference like this:
- Apple silicon: Ollama and Exo Labs. The card does not publish an Exo command, so this guide does not invent one.
- NVIDIA: SGLang, vLLM, Unsloth, NVIDIA Dynamo, and TokenSpeed, plus their dependency libraries.
- Quantized GGUF: Unsloth and llama.cpp.
- Also listed: Hugging Face, Llama App, and, for post-training, Miles, SkyRL, and NVIDIA NeMo AutoModel. Tinker is the lab's own inference and customization platform.
Weights and the research variable-loop checkpoint are on Hugging Face. Start with the download page if you want the file names, and with requirements if you are deciding whether a machine can hold the model.
Common questions
Is this Inkling 2 Mini guide official?
No. Thinking Machines Lab publishes the model, the announcement, and the model card. This site is an independent guide and is not affiliated with the lab.
What do 3.4B and 5.4B active parameters mean?
The model card says Inkling 2 Mini has 26.4 billion parameters and 3.4 billion active. Each extra pass through the looped layers adds about 1.0 billion effective active parameters, so three passes make 5.4 billion effective active parameters per token. Ollama's page quotes the 5.4 billion figure.
Can Inkling 2 Mini run on a laptop?
Thinking Machines says the checkpoint is about 20 GB and supports Apple silicon with 32 GB or more memory, one NVIDIA DGX Spark, and other GPUs with 32 GB or more of VRAM recommended. Smaller GGUF quants from Unsloth target less memory. See the requirements page before you buy hardware.
Sources
- Thinking Machines Lab, “Introducing Inkling 2 Mini” Announcement dated 20 November 2026.
- Inkling 2 Mini model card Hardware, license, architecture, evaluations, and limitations.
- Inkling 2 Mini config.json Hidden size, layer count, loop bounds, experts, and context length.
- thinkingmachines/Inkling-2-Mini on Hugging Face Official open weights.
- Hugging Face blog, “Welcome Inkling 2 Mini” Published 20 November 2026. Transformers, llama serve, SGLang, and MLX commands.
- Unsloth, Inkling 2 Mini run guide GGUF sizes, memory ranges, sampler settings, and llama.cpp commands.
- Ollama library page for inkling-2-mini
- Saad-Falcon, Narayan, et al., “Intelligence per Watt” arXiv:2511.07885, cited by the announcement.