Inkling by Thinking Machines Lab: Inside the New 975B-Parameter Open-Weights Multimodal Model
What Is Inkling?
On July 15, 2026, Thinking Machines Lab released Inkling, its first open-weights foundation model, alongside a preview of a smaller sibling, Inkling-Small. Inkling is a Mixture-of-Experts transformer with 975B total parameters and 41B active parameters, supporting a context window of up to 1M tokens, and was pretrained on 45 trillion tokens of text, images, audio, and video.
The release is notable less for topping leaderboards and more for its design philosophy. The company states plainly that Inkling is not the strongest overall model available today, open or closed — instead, its value lies in the combination of multimodal capability, efficient controllable thinking, and native availability on Tinker for fine-tuning. This is a bleeding-edge release explicitly optimized for customization, not leaderboard supremacy.
Architecture: What's Under the Hood
Inkling departs from the standard MoE playbook in several verifiable ways:
- MoE design: The design largely follows DeepSeek-V3, with each MoE layer containing 256 routed experts and 2 shared experts, and 6 routed experts active per token, using a sigmoid-based router with an auxiliary-loss-free load-balancing bias.
- Attention mechanism: The model interleaves sliding-window and global attention layers at a 5:1 ratio with 8 KV heads, and uses relative positional embeddings instead of the more commonly adopted Rotary Positional Embedding (RoPE) for better long-context extrapolation.
- Additional stabilization: Short convolutions are applied after key/value projections in each attention layer and on the attention/MLP residual branch outputs before they rejoin the main residual stream.
- Multimodal encoding: The model uses an encoder-free architecture, with audio input as dMel spectrograms and images encoded as 40x40 pixel patches via a four-layer hMLP.
Training at Scale
Inkling was trained with a hybrid optimizer — Muon for large matrix weights, Adam for other parameters — with weight decay strength coupled to the square of the learning rate to keep model weight magnitudes stable across the training horizon. Post-training was bootstrapped via SFT on synthetic data from open-weights models including Kimi K2.5, with the majority of compute spent on large-scale reinforcement learning across synthetic and human-created environments.
On the RL side: the company scaled reinforcement learning to over 30 million rollouts across two long continuous training runs, with reasoning performance improving log-linearly throughout. An unexpected side effect emerged during this process — the model's chain-of-thought became progressively more concise over training, dropping grammatical overhead while remaining comprehensible, an effect that was not directly targeted by the reward function but emerged from optimizing for efficiency alone.
Controllable Thinking Effort
A headline feature is token-efficient, adjustable reasoning depth. Inkling reaches a given benchmark score using fewer tokens than competitors — for instance, matching Nemotron 3 Ultra's Terminal Bench 2.1 score using roughly one-third the token count. This effort was trained explicitly: the team varied the model's effort level during RL rollouts by changing system messages and adjusting per-token cost, teaching the model to use different token budgets across different problems.
Benchmark Performance (Verified Numbers)
Selected results from the official model card, benchmarked at effort=0.99:
| Benchmark | Inkling | Nemotron 3 Ultra | Kimi K2.6 | GLM 5.2 | Gemini 3.1 Pro (high) | Claude Fable 5 (max) | GPT 5.6 Sol (max/xhigh) |
|---|---|---|---|---|---|---|---|
| HLE (text only) | 29.7% | 26.6% | 35.9% | 40.1% | 44.7% | 53.3% | 47.2% |
| SWEBench Verified | 77.6% | 70.7% | 80.2% | 80.0% | 80.6% | 95.0% | 82.2% |
| Terminal Bench 2.1 (Best Harness) | 63.8% | 56.4% | 71.3% | 82.7% | 73.8% | 84.6% | 89.5% |
| GPQA Diamond | 87.2% | 86.7% | 91.1% | 89.5% | 94.1% | 92.6% | 94.1% |
| StrongREJECT (safety) | 98.6% | 98.7% | 99.8% | 98.5% | 98.0% | 98.7% | 98.5% |
Inkling scored strongest relative to open-weights peers on FORTRESS Adversarial (78.0%), a benchmark testing refusal of harmful requests without over-refusing benign look-alikes.
Epistemics: Calibration, Instruction-Following, and Censorship Resistance
Thinking Machines put unusual emphasis on what it calls "epistemics." The model was trained via RL against proper scoring rules on a large corpus of resolved real-world forecasting questions to improve calibration. Instruction-following was trained using two automated graders — a rubric grader scoring responses against a checklist, and a claims grader that performs agentic web search to verify each factual claim and penalize unsupported ones, together improving helpfulness while reducing hallucination.
On censorship resistance specifically: the model was evaluated by Cognition on their Propaganda and Censorship Eval and exhibited strong patterns of censorship non-compliance.
Inkling-Small: The Efficient Sibling
Inkling-Small is a 276B-parameter MoE model with 12B active parameters (versus 41B for full Inkling), sharing the same post-training stack, and matches or exceeds Inkling on many benchmarks due to improvements made specifically to its pretraining data and recipe. It's currently in preview, with full weights pending completion of testing.
Availability and Fine-Tuning
Inkling is available on Tinker today with context length options of 64K and 256K tokens, offered at a 50% discount for a limited time. Full weights are published on Hugging Face, both as the original checkpoint and as an NVFP4 checkpoint optimized for NVIDIA Blackwell inference.
Inference partnerships extend the deployment surface significantly: the model is available via APIs on TogetherAI, Fireworks, Modal, Databricks, and Baseten, with additional open-source inference and RL support built with RadixArk for SGLang and Miles, Inferact for vLLM, Lightseek for TokenSpeed, Unsloth for llama.cpp, and Hugging Face for transformers integration.
Bottom Line
Inkling isn't pitched as the best model on any single leaderboard — its own release notes say so directly. Its real differentiation is architectural transparency (published MoE and attention design choices), genuine open weights (not just open API access), native multimodality trained from scratch, and a token-efficient controllable-reasoning mechanism that's directly verifiable via the effort-sweep benchmarks. For teams evaluating open-weights options to fine-tune on Tinker, the combination of full weight availability, broad inference-partner support, and documented calibration/epistemics training makes this a genuinely state-of-the-art open release as of July 2026 — provided the workload favors customization over raw peak benchmark scores.