Thinking Machines Unveils First Open-Weight Model Inkling
모델 출시 | Thu Jul 16 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 4 sources
Mira Murati's Thinking Machines Lab released Inkling, a 975B MoE multimodal open-weight model.
Analysis
[Thinking Machines Lab] released its first open-weight model Inkling [1][2][4]
- MoE architecture with 975B total parameters and 41B active
- Supports 1M token context window
- Pre-trained on 45 trillion tokens across text
- image
- audio
- and video
- Outputs limited to text (including code and structured data)
[Inkling] applied native multimodal reasoning and architectural innovations [3]
- Decoder-only multimodal MoE with 256 experts
- Uses relative attention instead of RoPE
- Hybrid attention alternating global and sliding window attention at a 5:1 ratio
- Short 1D convolution (SConv) for handling local representations
- Provides BF16 and NVFP4 variants and speculative MTP layers
[Inkling-Small] simultaneously released as a lightweight preview version [2]
- Reduced to 12B active parameters
- Applies the same training recipe
- Targets strong performance at lower cost and latency
[Tinker platform] supports Inkling fine-tuning and customization [1][2][4]
- Inkling fine-tuning immediately available on Tinker
- Added Inkling Playground console for developers
- Demonstrated Inkling writing
- running
- and evaluating its own fine-tuning jobs
- Customization safety is the user's responsibility
[Thinking Machines strategy] emphasized a decentralized, organization-tailored AI vision [1][4]
- Differentiated from the general-purpose chatbot approach of OpenAI
- Anthropic
- and Google
- Designed calibrated answers that indicate uncertainty
- Provides controls to adjust thinking effort
- Claims equivalent coding performance to Nvidia Nemotron 3 Ultra with 1/3 the tokens
- Startup valued at $12B seed valuation