Inkling
A 975B-parameter open-weights MoE multimodal model from Thinking Machines Lab that processes text, images, and audio through a 1M-token context window.
Get API key- Provider
- Thinking Machines Lab
- Price
- $2.34 in · $5.85 out / 1M
- Context window
- 1M tokens
- Released
- July 15, 2026
- License
- Apache 2.0
What is Inkling?
Inkling is a general-purpose multimodal transformer from Thinking Machines Lab, released July 15, 2026. It is a 66-layer sparse Mixture-of-Experts model with 975B total and 41B active parameters, accepting text, image, and audio inputs to generate text outputs across a 1M-token context window.
Use it privately on Venice
On Venice, Inkling runs privately with zero retention — your prompts are not stored, profiled, or used for training. You get the full open-weights model with vision, reasoning, tool use, and web search capabilities, quantized to FP4 for efficient inference without sacrificing the 1M context window. Because the weights are open under Apache 2.0, you can also download and self-host, but Venice lets you use it permissionlessly without infrastructure overhead.
What can it do?
- •Open weights released under Apache 2.0, enabling fine-tuning, self-hosting, and downstream integration without commercial restrictions.
- •Native multimodal reasoning over text, images, and audio through a shared decoder space.
- •Agentic coding and tool-use performance suitable for building autonomous systems and coding assistants.
- •Efficient sparse MoE design — 975B total parameters with only 41B active per token, lowering inference cost relative to dense frontier models.
- •Supports structured JSON output, multiple image inputs, and web search grounding when run on Venice.
- •Controllable thinking effort lets developers trade latency for depth on complex tasks.
- •Not the strongest overall model on every benchmark — Thinking Machines Lab explicitly notes it is a balanced foundation rather than a closed-frontier killer.
- •Very large total parameter count (975B) means self-hosting requires significant GPU infrastructure despite sparse activation.
- •Matched by some closed rivals on context length and exceeded in cost-efficiency by smaller open models for narrow tasks.
- •Audio and vision capabilities are frontier-competitive but may lag behind specialized closed multimodal models in specific niches.
- •FP4 quantization on Venice reduces memory but may slightly impact precision compared to BF16 or MXFP8 native runs.
Inkling capabilities
- Tool use / function calling
- Vision (image input)
- Reasoning
- Web search
- Code-optimized
- Structured output (JSON schema)
- Audio input
- Video input
- Multiple image inputs
- Log probabilities
How to use it via API
Venice exposes an OpenAI-compatible API. Swap your base URL and call inkling.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "inkling",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Specifications
Pricing
Billed per token on Venice: $2.34 per 1M input tokens and $5.85 per 1M output tokens.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
Inkling vs alternatives
| Model | Context window | Strongest at | Open weights | Price (Venice) |
|---|---|---|---|---|
| Inkling | 1M tokens | Multimodal open-weight generalist | Yes | $2.34 in · $5.85 out / 1M |
| Claude Fable 5 | 1M tokens | Long-context reasoning | No | $12 in · $60 out / 1M |
| DeepSeek V3.2 | 160K tokens | Cost-efficient coding | Yes | $0.33 in · $0.48 out / 1M |
| Kimi K2.6 | 256K tokens | Long-context open weights | Yes | $0.75 in · $3.50 out / 1M |
| Grok 4.5 | 500K tokens | Real-time web knowledge | No | $2.27 in · $6.80 out / 1M |
Balanced foundation with native vision, audio, and tool use under an open license.
What is it good for?
- •Building agentic AI systems that need tool use, reasoning, and multimodal perception.
- •Coding assistants and retrieval-augmented generation over long codebases.
- •Multimodal chatbots that accept screenshots, voice memos, and text in the same session.
- •Fine-tuning domain-specific variants on Tinker for enterprise or research use.
- •Long-context document analysis and synthesis.
Prompting tips
- •Use explicit multi-step instructions — Inkling is trained for detailed instruction following.
- •For coding tasks, specify the target language and desired output format such as JSON schema.
- •When using images, describe what to focus on if the task is specific; the hierarchical patch encoder handles general visual reasoning well.
- •Leverage controllable effort by asking for step-by-step reasoning when accuracy matters more than speed.
- •For audio inputs, ensure clear waveform quality; the model uses discrete token encoding.
Version history
Previewed alongside Inkling with 12B active parameters for lower-latency use.
CurrentCurrent release — 975B total / 41B active parameters, open weights under Apache 2.0.
Frequently asked questions
Inkling is a general-purpose multimodal transformer from Thinking Machines Lab, released in July 2026. It accepts text, image, and audio inputs and generates text outputs, using a 975B-parameter sparse MoE architecture with 41B active parameters and a 1M-token context window.
On Venice, Inkling is billed per token: $2.34 per 1M input tokens and $5.85 per 1M output tokens. There is no subscription required.
You can try Inkling on Venice with free credits included in new accounts. For continued use, you pay per token through Venice's credit system.
Yes. Inkling is released with open weights under the Apache 2.0 license. You can download the weights from Hugging Face, fine-tune it on Tinker, or self-host it. Venice provides permissionless API access without infrastructure setup.
Yes. Inkling supports function calling and tool use on Venice, making it suitable for agentic workflows that call external APIs or execute code.
Yes. Inkling is natively multimodal: it can process text, images, and audio inputs. On Venice it supports vision including multiple image inputs, audio input, and structured JSON output.
Inkling wins on openness, cost, and customizability: it is Apache 2.0 open-weights and far cheaper per token. Claude Fable 5 may edge out on raw reasoning benchmarks and also offers a 1M context window, but it is closed, proprietary, and roughly 5–10× more expensive on Venice.
Inkling supports up to 1,000,000 tokens of context, with a maximum output of 65,536 tokens per generation.
Yes. Thinking Machines Lab offers fine-tuning for Inkling on its Tinker platform. Because the weights are open, you can also fine-tune independently using your own infrastructure.
Related models
Run Inkling privately.
No prompt logging. No data used for training. Free to start — no credit card.
