Inkling 2 Small

Name: Inkling 2 Small
Author: thinkingmachines

by thinkingmachines

Thinking Machines' open multimodal MoE: images and text in, 1M-token context, 17B of 594B parameters active per token, Apache 2.0

274likes

594B total (17B active)params

HuggingFace Run it on your own data, free

Identifiers

Model ID

thinkingmachines/Inkling 2 Small

Deploy Inkling 2 Small

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.

Read the customer-hosted guide Talk to us about single-tenant

Overview

Inkling 2 Small is an open-weight multimodal model from Thinking Machines Lab, published on Hugging Face on 4 November 2026 under the Apache 2.0 license. It reads text and images, writes text, and takes up to one million tokens of context. It is built for agents, tool use, coding assistants and retrieval-augmented generation. The card reports 94.6% on OmniDocBench v1.6 for document parsing, 81.6% on MMMU Pro and 83.8% on ScreenSpot Pro for finding elements in screenshots. It does not take video directly, so video goes in as sampled frames.

Architecture

A 48-layer decoder-only transformer with a sparse Mixture-of-Experts feed-forward backbone: each token is routed to 8 of 384 experts, so 17B of its 594B parameters are active per token. Attention mixes local and global layers. Images are encoded by a vision transformer and processed together with text. It was trained in NVFP4, which is the main released checkpoint; an MXFP8 checkpoint is also published.

Mixpeek SDK Integration

# Serve it on your own GPUs (the card asks for at least 500 GB of VRAM for the
# NVFP4 checkpoint, e.g. 4x H200 or 2x B300). See the vLLM recipe for flags:
# https://recipes.vllm.ai/thinkingmachines/Inkling-2-Small
#   vllm serve thinkingmachines/Inkling-2-Small --tensor-parallel-size 4
import base64, requests
from openai import OpenAI

llm = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
img = base64.b64encode(open("page_4.png", "rb").read()).decode()
text = llm.chat.completions.create(
    model="thinkingmachines/Inkling-2-Small",
    messages=[{"role": "user", "content": [
        {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{img}"}},
        {"type": "text", "text": "Transcribe this page, including the numbers in any chart."},
    ]}],
).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": "reports/q3.pdf#page=4"}},
)

Capabilities

Image and text input, text output
Contexts up to 1M tokens
17B active parameters per token out of 594B
Document parsing, chart reading and screenshot grounding
Apache 2.0 license

Use Cases on Mixpeek

Describing page images, charts and screenshots as text you can index and search

Agents that read documents and screens and call tools

Answering questions over long document sets in one context

Writing structured fields from video frames for a search index

Benchmarks

Dataset	Metric	Score	Source
OmniDocBench v1.6	Document parsing	94.6%	Model card: thinkingmachines/Inkling-2-Small (self-reported)
MMMU Pro	Standard, 10 options	81.6%	Model card (self-reported)
ScreenSpot Pro	Screenshot grounding	83.8%	Model card (self-reported); 92.0% with Python
CharXiv RQ	Chart reasoning	85.6%	Model card (self-reported); 89.1% with Python
MuirBench	Multi-image understanding	79.2%	Model card (self-reported)

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 NVFP4 checkpoint needs at least 500 GB of GPU memory and the MXFP8 one at least 800 GB, per the card. Hosted APIs are listed on the card, including Together, Baseten, Modal and Cloudflare Workers AI.

Common Pipeline Companions

Qwen/Qwen3-VL-8B-Instruct

A small open vision-language model for high-volume frame and page descriptions

google/siglip2-base-patch16-224

Image-text embeddings to pick the frames or pages worth sending to the large model

Frequently Asked Questions

What is Inkling 2 Small?

An open-weight multimodal Mixture-of-Experts model from Thinking Machines Lab, published on 4 November 2026 under Apache 2.0. It reads text and images, writes text, takes up to one million tokens of context, and activates 17B of its 594B parameters per token.

Can Inkling 2 Small read images and documents?

Yes. It takes images as input alongside text. Thinking Machines reports 94.6% on OmniDocBench v1.6 for document parsing, 85.6% on CharXiv reasoning questions about charts, and 83.8% on ScreenSpot Pro for locating elements in screenshots.

Does Inkling 2 Small 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.

What hardware does Inkling 2 Small need?

The card asks for at least 500 GB of GPU memory for the main NVFP4 checkpoint, for example 2x NVIDIA B300 or 4x H200, and at least 800 GB for the MXFP8 checkpoint. It runs with SGLang, vLLM, TokenSpeed, Unsloth or Hugging Face Transformers, and several inference providers host it.

Can I search what Inkling 2 Small 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. Pair it with an image embedding model such as SigLIP 2 to search the images themselves.