Source: Databricks Blog — 2026-07-15
Summary
Thinking Machines Lab shipped Inkling, its first open-weights model, and Databricks announced same-day availability through Unity AI Gateway, making it a launch partner rather than a follow-on integration. Inkling is a 975B-parameter Mixture-of-Experts transformer with only 41B active parameters per token, natively handling text, images, and audio with a controllable "thinking effort" dial and up to a 1M-token context window. On Databricks, the model is governed like any other endpoint — centralized permissions, cost controls, and observability through Unity Catalog — and can be wired directly into coding agents such as Cursor and OpenCode.
Key Takeaways
- Inkling is a 66-layer decoder-only transformer with a sparse MoE feed-forward backbone: each token is routed to 6 of 256 experts plus 2 shared experts that fire on every token, yielding 975B total parameters but only 41B active per forward pass.
- It was pretrained on 45 trillion tokens spanning text, images, audio, and video, and reasons over audio natively via dMel spectrograms and over images as 40x40-pixel patches — unusual for an open-weights release, which are typically text-only or bolt on vision separately.
- "Controllable thinking effort" lets developers trade latency for quality along a single dial; Thinking Machines reports Inkling can match Nemotron 3 Ultra's Terminal Bench 2.1 score using roughly a third of the tokens when effort is tuned down.
- Reported benchmarks: 78.0% on FORTRESS Adversarial (leading the open-weights group), 73.5% on MMMU Pro, and 91.4% on VoiceBench — all at effort=0.99, temperature 1.0, with a 256K trajectory limit for coding evals.
- A lighter sibling, Inkling-Small (12B active parameters), ships alongside the flagship for lower-cost, lower-latency deployments using the same training recipe.
- On Databricks, Inkling is governed through Unity AI Gateway — the same layer that already brokers access to other models, agents, and MCP servers — so enterprises get centralized security, permissions, cost controls, and observability without a separate integration path.
Reel Script
Hook (~18s, 40 words): A brand-new 975-billion-parameter AI model just launched — and it only "wakes up" 41 billion parameters at a time to answer you. That's not a bug, that's the entire trick that makes trillion-parameter models affordable to actually run.
Core Concept (~70s, 160 words): This is Inkling, the first open-weights model from Thinking Machines Lab, and it's a Mixture-of-Experts model — picture a company with 256 specialists on staff, but for any single question, only 6 of them plus 2 generalists actually get pulled into the meeting. That's why "975B parameters" is misleading on its own — the real cost driver is the 41B that fire per token. The more interesting part is what Thinking Machines calls "controllable thinking effort": a dial you can turn to make the model think longer and harder, or answer fast and cheap, for the same underlying weights. They claim that at a tuned-down effort setting, Inkling matches a rival model's coding benchmark score while burning roughly a third of the tokens. And unlike most open-weights releases that are text-only, Inkling natively reads audio and images — no bolted-on vision encoder, it's baked into the architecture from pretraining.
Hands-On (~110s, 250 words): Look at how the architecture is actually described: 66 decoder layers, sparse MoE feed-forward blocks, 256 experts per layer, top-6 routing plus 2 always-on shared experts. That routing decision — which 6 of 256 experts handle this specific token — happens fresh for every single token in the sequence, which is the mechanism worth sketching on a whiteboard: input token comes in, a small router network scores all 256 experts, top 6 get activated, their outputs get combined, next token repeats the whole process. For multimodal input, images get chopped into 40x40-pixel patches before entering that same transformer stack, and audio gets converted into dMel spectrograms — both modalities riding the same MoE backbone rather than separate specialized towers. On the deployment side, Databricks slots Inkling behind Unity AI Gateway, the same governance layer that already fronts other models and MCP servers — meaning a security team gets one place to set per-model cost caps, permission scopes, and usage logs, instead of bolting on a separate access-control layer just because this particular model came from a different lab. That's the part enterprise viewers should care about even if they never touch the model weights directly: governance now travels with the model, not around it.
Takeaway (~25s, 55 words): The interesting story here isn't the parameter count — trillion-parameter open models aren't new — it's that "controllable effort" gives you one model instead of three separate size tiers. Worth trying if you're evaluating open-weights options for a cost-sensitive coding agent. Go read the model card before you assume it's a straight Kimi or Nemotron swap.