NEWVectors or files. Pick a path.Start →
    Models/thinkingmachines/Inkling 2 Mini
    Apache 2.0

    Inkling 2 Mini

    by thinkingmachines

    Thinking Machines' small open multimodal model: images and text in, a 1M-token context, 3.4B of 26.4B parameters active, and a 20 GB checkpoint that runs on a 32 GB Mac

    1,064likes
    26.4B total (3.4B active)params
    Identifiers
    Model ID
    thinkingmachines/Inkling 2 Mini

    Deploy Inkling 2 Mini

    Single-tenant

    Mixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.

    Overview

    Inkling 2 Mini is an open-weight multimodal model from Thinking Machines Lab, dated November 20, 2026 on its model card and released under the Apache 2.0 license. It reads text and images, writes text, and takes up to one million tokens of context. It is the smaller sibling of Inkling 2 Small, built to run on local hardware: the single checkpoint is about 20 GB in mixed NVFP4 and MXFP8 precision, and the card lists an NVIDIA DGX Spark or an Apple silicon machine with 32 GB of memory as supported.

    It is the first looped transformer Thinking Machines has released. Its middle 12 layers run three times with shared weights, so each token passes through 48 layer applications from 24 layers of weights. Each layer routes a token to 8 of 128 experts, which gives 3.4B active parameters out of 26.4B, or about 5.4B effective active parameters once the three passes are counted.

    On the card's vision benchmarks it scores 92.8% on OmniDocBench v1.6 for document parsing and 74.9% on MMMU Pro. For a media or document pipeline, it can turn pages, screenshots and sampled video frames into text you can index and search, on hardware a single team can own.

    Architecture

    A 24-layer decoder-only transformer with sparse Mixture-of-Experts feed-forward layers, routing each token to 8 of 128 experts. The middle 12 layers are looped: they run three times with shared weights, for 48 layer applications per token. Attention is a hybrid of local and global layers. Images are encoded by a vision transformer and projected into the same hidden space as text, and the decoder processes both together.

    Mixpeek SDK Integration

    # Serve it on one GPU (the card recommends 32 GB or more of VRAM). The vLLM recipe
    # needs vLLM 0.35.0 or the vllm/vllm-openai:inkling-2-mini image:
    # https://recipes.vllm.ai/thinkingmachines/Inkling-2-Mini
    #   vllm serve thinkingmachines/Inkling-2-Mini --tensor-parallel-size 1 --reasoning-parser inkling
    import base64, requests
    from openai import OpenAI
    
    llm = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
    img = base64.b64encode(open("frame_0042.jpg", "rb").read()).decode()
    text = llm.chat.completions.create(
        model="thinkingmachines/Inkling-2-Mini",
        messages=[{"role": "user", "content": [
            {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{img}"}},
            {"type": "text", "text": "Describe what is happening in this frame in one sentence."},
        ]}],
        max_tokens=400,
    ).choices[0].message.content
    
    # Store the description as a text object so a Mixpeek text collection can search it.
    requests.post(
        "https://api.mixpeek.com/v1/buckets/bkt_your_bucket/objects",
        headers={"Authorization": "Bearer API_KEY", "X-Namespace": "ns_your_namespace"},
        json={"blobs": [{"property": "description", "type": "text", "data": text}],
              "metadata": {"source_key": "clips/warehouse.mp4#t=42"}},
    )

    Capabilities

    • Image and text input, text output
    • Contexts up to 1M tokens
    • 3.4B active parameters per token out of 26.4B (about 5.4B effective with the three looped passes)
    • A single 20 GB checkpoint in NVFP4 and MXFP8 that runs on one GPU or a 32 GB Mac
    • Apache 2.0 license

    Use Cases on Mixpeek

    Describing page images, charts and screenshots as text you can index and search
    Writing captions for sampled video frames on a single GPU or workstation
    Reading documents and screens inside a local agent that calls tools
    A first pass over large image sets before sending hard cases to a bigger model

    Benchmarks

    DatasetMetricScoreSource
    OmniDocBench v1.6Document parsing92.8%Model card: thinkingmachines/Inkling-2-Mini (self-reported; Qwen3.8-27B 91.5%)
    MMMU ProStandard, 10 options74.9%Model card (self-reported; Qwen3.8-27B 76.3%)
    ScreenSpot ProScreenshot grounding76.5%Model card (self-reported); 87.7% with Python
    MuirBenchMulti-image understanding77.5%Model card (self-reported)
    PerceptionBenchVisual perception57.5%Model card (self-reported; Qwen3.8-27B 52.4%)

    Performance

    Input SizeText and images (ideally 40 to 4,096 px per side), up to 1M tokens of context
    Embedding Dimn/a (outputs text)
    GPU LatencyInput dependent
    GPU ThroughputBatch dependent
    GPU MemoryModel dependent

    We have not measured it. The scores are Thinking Machines' own. The checkpoint is about 20 GB, and the card recommends 32 GB or more of GPU memory or an Apple silicon machine with 32 GB. Because the middle layers run three times, compute per token is higher than its 3.4B active parameters suggest, and the card recommends limits on output tokens because it can loop in long reasoning.

    Frequently Asked Questions

    What is Inkling 2 Mini?

    An open-weight multimodal model from Thinking Machines Lab under Apache 2.0, dated November 20, 2026 on its card. It reads text and images, writes text, takes up to one million tokens of context, and activates 3.4B of its 26.4B parameters per token. It is a smaller, local-hardware version of Inkling 2 Small.

    Can Inkling 2 Mini run on a laptop?

    On a large one. The checkpoint is about 20 GB in mixed NVFP4 and MXFP8 precision, and the card lists Apple silicon MacBooks, Mac minis and Mac Studios with 32 GB or more of memory, an NVIDIA DGX Spark, or any GPU with 32 GB or more of VRAM. Ollama and Exo Labs run it on Apple silicon, and Unsloth publishes quantized GGUF versions for llama.cpp.

    What is a looped transformer?

    A transformer that runs some of its layers more than once with the same weights. Inkling 2 Mini runs its middle 12 of 24 layers three times, so each token passes through 48 layer applications. That adds computation without adding parameters: each extra pass adds about 1.0B effective active parameters, so the model behaves like about 5.4B active parameters per token while storing 26.4B in total.

    Does Inkling 2 Mini understand video?

    Not directly. It accepts text and images, so a video is sent as sampled frames, for example one per scene or one every few seconds, with their timestamps in the prompt. Its training data included video, according to the card.

    Can I search what Inkling 2 Mini reads with Mixpeek?

    Yes. Store each description it writes as a text object with the key of the page, image or frame it came from, and a Mixpeek text collection and retriever make those descriptions searchable alongside any image embeddings in the same namespace.

    Specification

    Organizationthinkingmachines
    Retriever-
    Parameters26.4B total (3.4B active)
    LicenseApache 2.0
    Downloads/moN/A
    Likes1,064

    Build a pipeline with Inkling 2 Mini

    Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.

    Run it on your own data, free