Inkling Model Card
5 hours ago
- #multimodal-ai
- #open-weights
- #safety-evaluation
- Inkling is a multimodal autoregressive transformer model with 975B total parameters (41B active) released by Thinking Machines Lab on July 15, 2026, under Apache 2.0 license.
- It accepts text, image (40px-4096px), and audio (16kHz WAV up to 20 mins) inputs, generating text outputs in English and multiple languages, with a 1M token context window.
- Intended for developers building AI applications like chatbots, coding assistants, and agentic systems, it's available via API (Tinker/third-party) and open weights on Hugging Face.
- Hardware requirements: BF16 checkpoint needs 2TB VRAM (e.g., 8x NVIDIA B300), while NVFP4 quantized version requires 600GB VRAM (e.g., 4x NVIDIA B300 with SM100+).
- Training data includes public, third-party, and synthetic sources, curated via deduplication and filtering for quality and safety, but may contain biases from web-derived data.
- Evaluations show strong performance in reasoning (e.g., 46.0% HLE with tools), agentic tasks (e.g., 77.6% SWEBench Verified), and multimodal benchmarks like vision (73.5% MMMU Pro).
- Safety assessments indicate risks are consistent with open-weight models, with residual issues in role-play compliance; recommendations include using tools like Llama Guard for moderation.
- Limitations include hallucination, instruction-following failures, and bias; knowledge is cut off at training date, so verification is advised for high-stakes uses like medical or legal domains.