Inkling: A 975B-parameter multimodal MoE model with open weights and controllable reasonin
A 975-billion-parameter open-weights model that lets you dial up or down reasoning effort on the fly has just landed from Thinking Machines Lab. Named Inkling, the new Mixture-of-Experts transformer is positioned as a customizable base for developers who want to adapt it to niche tasks without starting from scratch. Weights are fully open and fine-tunable on the lab’s Tinker platform, while the architecture mixes a 66-layer decoder with a sparse MoE feed-forward backbone that activates only six routed experts plus two shared experts per token.
Built for multimodal inputs—and a million-token window Inkling ingests text, images, and audio (as dMel spectrograms) and outputs UTF-8 text only. Images are split into 40×40 patches processed by a four-layer hMLP encoder, while audio is tokenized via spectrograms before being projected into the same embedding space as text. Positional encoding uses a relative scheme rather than RoPE, a choice the team says improves extrapolation. The context window stretches to a million tokens, a practical boon for long-document or extended-conversation scenarios.
Reasoning cost becomes a tunable knob
Unlike traditional models whose compute profile is fixed, Inkling exposes a controllable “thinking effort” that scales from 0.2 to 0.99. During reinforcement learning the team varied effort via system prompts and per-token cost, teaching the model to budget tokens differently across rollouts. In practice, developers can set an effort argument (with named levels) or use the reasoning_effort parameter to trade latency for quality on the fly. Efficiency numbers suggest real savings: Inkling hits the same Terminal Bench 2.1 score as Nemotron 3 Ultra while using roughly one third the tokens, according to the lab’s figures.
Early external benchmarks show Inkling competitive with proprietary and open peers at maximum effort. On HLE (text only) it scores 29.7%, compared with 40.1% for DeepSeek V4 Pro. It tops 97% on AIME 2026 and 87% on GPQA Diamond, while SWE-bench Verified lands at 77.6%. Terminal Bench 2.1 sits at 63.8%, MCP Atlas at 74.1%, and SimpleQA Verified at 43.9%. The team also previewed Inkling-Small (276B total / 12B active), which matches or exceeds its larger sibling on many of these tests once weights are released.
Why it matters
Inkling’s open-weight release lowers the barrier to domain-specific fine-tuning and broadens access to cutting-edge multimodal capabilities. The controllable reasoning effort turns inference into an adjustable dial rather than a fixed profile, giving practitioners a practical lever to balance cost, latency, and accuracy. For teams building specialized assistants or pipelines, the combination of open weights, a million-token window, and tunable compute could shift the economics of deploying frontier models.
Source: MarkTechPost. AI-assisted editorial synthesis — TechnoExpress.