Inkling-Small

Inkling-Small

Thinking Machines Lab
+
+

Related Products

  • Gemini Enterprise Agent Platform
    1,161 Ratings
    Visit Website
  • TrustInSoft Analyzer
    6 Ratings
    Visit Website
  • LTX
    182 Ratings
    Visit Website
  • Checksum.ai
    2 Ratings
    Visit Website
  • Buildium
    2,562 Ratings
    Visit Website
  • RetailEdge
    201 Ratings
    Visit Website
  • TIMi
    68 Ratings
    Visit Website
  • Acadia
    3 Ratings
    Visit Website
  • Fraud.net
    56 Ratings
    Visit Website
  • rezStream Cloud PMS
    126 Ratings
    Visit Website

About

Inkling-Small is an efficient model that offers performance comparable to Inkling at a quarter of its size. It is a Mixture-of-Experts transformer with 276 billion total parameters and 12 billion active parameters, trained on NVIDIA GB300 NVL72 systems. It supports native reasoning across text, images, and audio, variable thinking effort, and context windows of up to one million tokens. Users adjust reasoning effort from minimal to extra high to balance performance and compute. Improved pre-training data, post-training with on-policy distillation from Inkling, and extended agentic coding reinforcement learning helped Inkling-Small surpass its larger counterpart on reasoning and coding benchmarks. It performs well in coding and tool-use harnesses, exceeds 80% on SWE-bench Verified, and combines strong reasoning with efficient output. Its encoder-free multimodal architecture processes audio as dMel spectrograms and images as 40-by-40-pixel patches alongside text tokens.

About

Muse Spark Open is an open-weight, natively multimodal language model for developers and organizations building AI applications, coding systems, and agentic workflows. The released model has 816 billion total parameters, 42 billion active parameters, and a 1-million-token context window. Teams can download the model weights, run inference through supported engines such as vLLM, SGLang, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, or Unsloth, and customize the model using PyTorch. It supports long-context reasoning, software engineering, coding, tool use, and multimodal inputs. Deployment options include cloud inference through listed API providers or self-hosting on compatible GPU systems; Meta documents a desktop setup using two Apple M5 Ultra Mac Studio computers. The model is released under the Muse Model License with limited commercial terms.

Platforms Supported

Windows Not Supported
Mac Not Supported
Linux Not Supported
Cloud Supported
On-Premises Not Supported
iPhone Not Supported
iPad Not Supported
Android Not Supported
Chromebook Not Supported

Platforms Supported

Windows Supported
Mac Supported
Linux Supported
Cloud Supported
On-Premises Supported
iPhone Not Supported
iPad Not Supported
Android Not Supported
Chromebook Not Supported

Audience

AI researchers and developers seeking an efficient open-weight multimodal model for reasoning, agentic tasks, and lower-cost deployment

Audience

AI developers, machine learning engineers, software engineers, AI researchers, enterprise development teams, and organizations seeking customizable open-weight multimodal models for local inference, autonomous agents, complex coding, and long-context AI applications

Support

Phone Support Not Supported
24/7 Live Support Not Supported
Online Supported

Support

Phone Support Not Supported
24/7 Live Support Not Supported
Online Not Supported

API

Offers API Supported

API

Offers API Not Supported

Screenshots and Videos

Screenshots and Videos

Pricing

$0.30 per million input tokens
$0.30 per million input tokens and $1.20 per million output tokens
Free Version Not Supported
Free Trial Not Supported

Pricing

$0
Model weights are available at no charge under the Muse Model License, which has limited commercial-use terms. Hosted inference providers set their own usage prices.
Free Version Supported
Free Trial Not Supported

Reviews/Ratings

Overall 5.0 / 5
features 4.0 / 5

Reviews/Ratings

Overall 5.0 / 5
ease 5.0 / 5

Pros & Cons from Real Users

Pros

  • Inkling-Small is really interesting from a developer’s point of view because it hits a sweet spot between serious model capability and practical deployability. A 276B-parameter model with only 12B active parameters per token is exactly the kind of architecture that makes sense if you care about cost, speed, and scaling real AI workflows. I also like that it is open weights under Apache 2.0. That makes it way more appealing for developers who want to fine-tune, inspect, customize, or build on top of the model without being completely locked into a closed API. The multimodal support is a big plus too. Being able to work with text, images, and audio inputs gives Inkling-Small a lot of room for developer tools, coding agents, support bots, document workflows, and internal automation.

Cons

  • The main downside is that “small” here is still not tiny. Even with only 12B active parameters, this is still a large open model that will require real infrastructure if you want to host it yourself. I would also want to test it deeply before making it the backbone of a production coding agent. The model card and early coverage look promising, but real developer workflows expose problems that benchmarks do not always catch: messy repos, flaky tests, weird dependencies, tool failures, and long multi-step tasks.

Pros & Cons from Real Users

Pros

  • Meta finally opened up a genuinely frontier-level Muse model. The 1M-token context window and 42B active parameters make it strong enough for large repos, long agent sessions, multimodal work, and complex coding without feeling like a stripped-down open model. The local deployment story is also a huge deal. Being able to run something this capable on two M5 Ultra Mac Studios is impressive, especially for teams that want more control over data, infrastructure, and customization. I also like how much of the stack is open around it. Muse Spark Open works with MLX, vLLM, SGLang, TensorRT-LLM, Transformers, llama.cpp, and Unsloth, and Meta is making the weights available through Hugging Face, SageMaker, Google Cloud Model Garden, and Microsoft Foundry.

Cons

  • Even though it is technically desktop-capable, needing two M5 Ultra Mac Studios for the full local experience is still a serious setup, so most developers will probably use hosted inference instead.

Training

Documentation Supported
Webinars Not Supported
Live Online Not Supported
In Person Not Supported

Training

Documentation Supported
Webinars Not Supported
Live Online Not Supported
In Person Not Supported

Company Information

Thinking Machines Lab
Founded: 2025
United States
thinkingmachines.ai/news/inkling-small/

Company Information

Meta
Founded: 2004
United States
research.meta.ai/blog/muse-spark-open

Alternatives

Alternatives

Inkling

Inkling

Thinking Machines Lab
Beam

Beam

Reflection
Laguna XS 2.1

Laguna XS 2.1

Poolside

Categories

AI Coding Models Supported
AI Models Supported
AI Vision Models Supported

Categories

AI Coding Models Supported
AI Models Supported
AI Receptionists Supported
Foundation Models Supported
Multimodal Models Supported

Integrations

Baseten Not Supported
Fireworks AI Not Supported
Groq Not Supported
Hugging Face Not Supported
Modal Not Supported
Model Context Protocol (MCP) Supported
Nebius Not Supported
OpenRouter Not Supported
Tinker Supported
Together AI Not Supported
vLLM Not Supported

Integrations

Baseten Supported
Fireworks AI Supported
Groq Supported
Hugging Face Supported
Modal Supported
Model Context Protocol (MCP) Not Supported
Nebius Supported
OpenRouter Supported
Tinker Not Supported
Together AI Supported
vLLM Supported
Claim Inkling-Small and update features and information
Claim Inkling-Small and update features and information
Claim Muse Spark Open and update features and information
Claim Muse Spark Open and update features and information