Skip to content
Computer vision

Inkling 2 Small (MXFP8)

a 594B multimodal MoE model with a 1M context window

2 min readupdated

Inkling 2 Small is a multimodal model from Thinking Machines that accepts text and images (from 40px to 4096px) to generate text responses. Built on a 48-layer transformer with an MoE architecture (Mixture-of-Experts, a sparse setup where 8 out of 384 modules activate per token to speed up inference), only 17B of its 594B total parameters run at any given time. The model features a 1M-token context window and hybrid attention combining local and global layers. Stored in the MXFP8 format, these weights are geared toward developers building AI agents, retrieval-augmented generation (RAG) systems, and coding assistants.

Key features

  • 594B / 17B MoE architecture: 48 decoder layers routing 8 of 384 experts per token, paired with a built-in vision encoder for processing text and images together.
  • Up to 1M-token context: Holds massive volumes of text and visual data in memory within a single request.
  • MXFP8 format: An 8-bit floating-point variant that scores 1–2 percentage points higher on benchmarks compared to the base NVFP4 release.
  • Broad framework support: Works out of the box for local inference via SGLang, vLLM, TokenSpeed, and Unsloth, as well as fine-tuning with Tinker, Miles, and SkyRL.

What you can do

  • Build autonomous AI agents: Power systems that rely on external tool-use and complex instruction following.
  • Run a coding assistant: Generate, analyze, and debug code across multiple programming languages while keeping large codebases in context.
  • Create multimodal RAG services: Analyze long documents, diagrams, and images inside a context window of up to 1M tokens.

Example

You can test the model in the interactive Tinker Playground, call it via API providers (NVIDIA, Fireworks, Together, Baseten, Nebius, Modal, Cloudflare), or deploy it locally using recipes for vLLM, SGLang, TokenSpeed, and Unsloth—passing in UTF-8 text alongside images ranging from 40×40 to 4096×4096 pixels.

Read moreDetails
Benchmarks: AA Index (v5.0) — 30.6, Vals Index (v2.2) — 56.1%, FrontierCode v2.0 — 25.4%.
Hardware requirements: The MXFP8 version requires a GPU cluster with at least 800 GB of total VRAM (such as 4× NVIDIA B300 or 8× NVIDIA H200).
Authors' note: The model was trained with NVFP4 as its base precision (which requires 500+ GB VRAM), and the authors consider MXFP8's 1–2 percentage point benchmark gain negligible for most practical workloads.
Discuss Telegram 0 24

Will it run?

Won't run

Not enough memory even for 4-bit. Below are GPUs you can rent.

  • F8_E4M3recommendedweights 560 GB · ready files616 GB
  • 4-bitweights 311 GB343 GB

Tip: go 8-bit.F8_E4M3 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

No single instance holds the whole model: it needs RAM offload or a cluster.

Your own GPU: Won't fit any home GPU — renting is cheaper.

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # omegazug/Inkling-2-Small-MXFP8
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download omegazug/Inkling-2-Small-MXFP8 --local-dir ~/models/Inkling-2-Small-MXFP8
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Inkling-2-Small-MXFP8
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of Inkling-2-Small-MXFP8 (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

omegazug/Inkling-2-Small-MXFP8
Author
omegazug
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

240

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments