Thinking Machines Inkling
What is Thinking Machines Inkling?
Inkling is an open-weight multimodal foundation model family from Thinking Machines Lab, Inc., released under the Apache 2.0 license. The models accept text, image, and audio inputs natively and generate text outputs. Intended use cases include agentic and tool-use systems, coding assistants, chatbots, retrieval-augmented generation (RAG) applications, and general conversational or instruction-following tasks.
Architecture & Technical Specifications
- Model Sizes & Active Parameters:
- Inkling: 975 billion total parameters, 41 billion active parameters per token.
- Inkling 2 Small: 594 billion total parameters, 17 billion active parameters per token.
- Inkling-Small: 276 billion total parameters, 12 billion active parameters per token.
- Backbone Architecture: 66-layer decoder-only transformer with a sparse Mixture-of-Experts (MoE) feed-forward backbone. Each token is routed to 6 of 256 routed experts, plus 2 shared experts active on every token. Attention uses a hybrid of local and global layers.
- Multimodal Processing: Images are processed via a hierarchical patch encoder, while audio inputs use discrete token encoding. All input modalities are projected into a shared hidden space and processed jointly within the decoder.
- Context Window: Up to 1,000,000 (1M) tokens.
- Numeric Precision Formats: BF16, MXFP8, and NVFP4 checkpoints.
- Input / Output Specifications:
- Text Input: UTF-8 encoded text.
- Image Input: Pixel-based formats (optimal performance when long and short dimensions range between 40px and 4,096px).
- Audio Input: 16kHz WAV format (optimal performance within 20 minutes of audio).
- Output: UTF-8 encoded text.
Infrastructure & Deployment Requirements
- Distribution Options: Downloadable open weights on Hugging Face; cloud API access via Thinking Machines’ Tinker platform, third-party inference providers, and hosted NVIDIA NIM microservices.
- Hardware Requirements (Self-Hosting):
- BF16 Checkpoint: Requires a GPU cluster with at least 2 TB of aggregated VRAM (e.g., 8× NVIDIA B300 or 16× NVIDIA H200 GPUs).
- NVFP4 Checkpoint: Requires at least 600 GB of aggregated VRAM (runnable as W4A4 on 4× NVIDIA B300 GPUs requiring SM100+ architecture, or W4A16 on 8× NVIDIA H200 GPUs).
- Supported Software Engines: SGLang, vLLM, TokenSpeed, Unsloth, or Hugging Face.
Vendor-Reported Evaluations (effort=0.99)
- Reasoning: AIME 2026: 97.1%; GPQA Diamond: 87.2%; HLE (text only): 29.7%; HLE (with tools): 46.0%.
- Agentic & Coding: SWE-bench Verified: 77.6%; SWE-bench Pro Public: 54.3%; Terminal Bench 2.1 (Best Harness): 63.8%; MCP Atlas: 76.0%; Tau 3 Banking: 23.7%; Toolathlon Verified: 45.5%; BrowseComp (with context management): 77.1%.
- Factuality & Chat: SimpleQA Verified: 43.9%; Global-MMLU-Lite: 88.7%; IFBench: 79.8%.
- Multimodal (Vision & Audio): MMMU Pro (Standard 10): 73.5%; CharXiv RQ: 78.1% (82.0% with Python harness); MMAU: 77.2%; VoiceBench: 91.4%.
- Safety: StrongREJECT: 98.6%; FORTRESS Benign: 95.9%; FORTRESS Adversarial: 78.0%.
Safety & Operational Limitations
Thinking Machines Lab notes inherent foundation model limitations, including hallucination risk, degraded instruction adherence over extended multi-turn sessions, demographic or domain bias, and knowledge boundaries capped at the training cutoff. Pre-release red-teaming evaluated sycophancy, manipulation, and dangerous capabilities (CBRN/cyber uplift), concluding Inkling poses no material uplift over existing open-weight ecosystems. The model card recommends defense-in-depth strategies (such as layering Llama Guard or external content moderation) and explicitly advises against unverified deployment in high-stakes medical, legal, or safety-critical contexts.
Categories & Use Cases
Technical Details
| Mobile Application | No |
|---|