Inkling
- Score
- 12.5 · rank 35
- Price
- Free
- API
- Yes
- Weights
- Yes
WHAT IT DOES
Inkling is the first open-weights model from Thinking Machines Lab, released 15 July 2026 under Apache 2.0. It is a 975B-parameter mixture of experts with 41B active per token, trained from scratch on 45 trillion tokens of text, images, audio and video, with a context window up to 1M tokens. The architecture uses 256 routed experts plus 2 shared experts per layer with 6 routed experts active per token, a sigmoid router, and interleaved sliding-window and global attention at a 5:1 ratio. It accepts text, image and audio input natively and returns text. Weights are on Hugging Face in both the original format and an NVFP4 checkpoint for Blackwell hardware. A distilled Inkling-Small followed on 31 July.
WHY WE PICKED IT
It is the first roughly trillion-parameter open-weights model that takes audio and images natively rather than through a bolted-on encoder, and Apache 2.0 means the weights can be used commercially without asking anyone. Thinking Machines is candid that this is not the strongest model available but a base worth customising, which is a more honest pitch than most open releases make.
QUICK FACTS
STRENGTHS
- ✓Apache 2.0, the least restrictive licence of any model at this scale
- ✓Native text, image and audio input in a single open checkpoint
- ✓97.1 percent on AIME 2026 and 77.6 percent on SWE-bench Verified in Thinking Machines' own evaluations
- ✓Runs on vLLM, SGLang, llama.cpp and transformers from day one
- ✓NVFP4 checkpoint published alongside the original weights
LIMITATIONS
- ⚠Distributed as code and model weights; there is no hosted product to sign into, so running it means self-hosting or a third-party host.
- ⚠Thinking Machines states plainly that it is not the strongest model available, it is a base for customisation
- ⚠The headline benchmark figures are vendor-run; on Artificial Analysis' independent composite it sits at 40.7, rank 36, well below the frontier
- ⚠975B parameters means serious hardware even at 41B active, so most teams will use a hosted provider
- ⚠Text output only, despite accepting three input modalities
GETTING STARTED
Pull the weights from huggingface.co/thinkingmachines/Inkling, or skip the hardware and call it through Together AI, Fireworks, Modal, Databricks or Baseten. For local work, the NVFP4 checkpoint is the one to use on Blackwell cards; everyone else should start from the original checkpoint on vLLM or SGLang. If the plan is to fine-tune rather than use it as shipped, that is what Thinking Machines built the Tinker platform for.
BENCHMARKS
QUICK TIPS
OFFICIAL LINKS
SIMILAR TOOLS
USE CASE EXAMPLES
Domain Fine-Tune On An Open Base
Adapt a frontier-scale multimodal model to a specific domain without licence negotiation.
- Confirm Apache 2.0 covers the intended commercial use
- Assemble a domain dataset spanning the modalities you actually need
- Fine-tune on Tinker or your own infrastructure
- Evaluate against the base checkpoint on your own tasks
- Deploy through vLLM or SGLang
- Keep the base model available as a control for regressions
Audio And Document Understanding In One Model
Process recordings and scanned pages without a separate speech pipeline.
- Convert recordings to 16kHz WAV segments under 20 minutes
- Send audio and page images in the same request
- Ask for a structured summary rather than free text
- Spot-check transcription-dependent claims against the source
- Batch through a hosted provider if volume is bursty
- Compare cost against a separate ASR plus LLM pipeline
PRICING
FEATURED IN GUIDES
How to Evaluate Inkling 2 Small on Your Own Code
A practical first evaluation for Inkling 2 Small: choose a bounded code change, define checks, revie...
When to Download Mistral Large 4
Mistral published the weights for Large 4, nicknamed Le Chonk, on October 29. What the licence allow...
Who Can Use Gemini 4 Carbon
Google announced Gemini 4 Carbon on October 28 for a small set of cyber defenders. Who has access, w...
When to Use GPT-6.1 Terra
OpenAI priced GPT-6.1 Terra at $1 per million input tokens for computer use and everyday writing. Ho...
Frontier Training Without the Live Web
OpenAI says frontier training will use cached and simulated web content, after a model reached the l...
EXPLORE ALTERNATIVES
Compare Inkling with 5+ similar llms AI tools.
FREQUENTLY ASKED QUESTIONS
Is Inkling free?
Yes, Inkling is completely free to use.
Does Inkling have an API?
Yes, Inkling offers an API for programmatic integration.
What is Inkling best for?
Inkling is best for Best Open-Weight Multimodal Base. Inkling is the first open-weights model from Thinking Machines Lab, released 15 July 2026 under Apache 2. It is the first roughly trillion-parameter open-weights model that takes audio and images natively rather than through a bolted-on encoder, and Apache 2.0 means the weights can be used commercially without asking anyone. Thinking Machines is candid that this is not the strongest model available but a base worth customising, which is a more honest pitch than most open releases make.
What platforms does Inkling support?
Inkling supports api, local.
Is Inkling open source?
Yes, Inkling is open source. You can access the source code, contribute, and deploy it on your own infrastructure.
How do I get started with Inkling?
Pull the weights from huggingface.co/thinkingmachines/Inkling, or skip the hardware and call it through Together AI, Fireworks, Modal, Databricks or Baseten. For local work, the NVFP4 checkpoint is the one to use on Blackwell cards; everyone else should start from the original checkpoint on vLLM or ...
How do I use Inkling?
Inkling is a large language model for text generation, analysis, and conversation. Use the API for programmatic access. Enter prompts or questions to get responses. It excels at apache 2.0, the least restrictive licence of any model at this scale.
FEATURED ON CURATEDAI
Work on Inkling? You're hand-reviewed in our directory. Add this badge to your site — it links back to this profile.
