An open-weights multimodal model by Thinking Machines Lab — 975B total parameters, 41B active, built for customization, agentic coding, and controllable thinking.
Chat with Inkling through the Hugging Face Inference API. Powered by Thinking Machines Lab's open-weights model.
Demo by akhaliq — uses the HF Inference Providers router.
Inkling is a broad generalist foundation model trained for real-world customization.
Adjust effort from 0.2 to 0.99 to balance performance vs. token cost. Reaches the same score as Nemotron 3 Ultra on Terminal Bench 2.1 at ~1/3 the tokens.
Scores 77.6% on SWE-Bench Verified and 63.8% on Terminal Bench 2.1. Builds web apps, games, and PDFs through tool use and long refinement loops.
Native multimodal reasoning over images, charts, diagrams, and audio. Encoder-free architecture using dMel spectrograms and hMLP patches.
Trained with proper scoring rules for calibrated confidence. Strong instruction following via rubric + claims graders with agentic web search.
Among the strongest open-weights safeguards on FORTRESS (78% adversarial) and StrongREJECT (98.6%). Commissioned external safety testers.
Available on Tinker for fine-tuning. Inkling can even self-fine-tune — it wrote and ran its own lipogram fine-tuning job and switched weights.
Scored at effort=0.99 across reasoning, coding, agentic, factuality, and multimodal tasks.
| Benchmark | Inkling | Nemotron 3 Ultra | Kimi K2.6 | GLM 5.2 |
|---|---|---|---|---|
| HLE (text only) | 29.7% | 26.6% | 35.9% | 40.1% |
| HLE (with tools) | 46.0% | 37.4% | 54.0% | 54.7% |
| AIME 2026 | 97.1% | 94.2% | 96.4% | 99.2% |
| GPQA Diamond | 87.2% | 86.7% | 91.1% | 89.5% |
| Benchmark | Inkling | Nemotron 3 Ultra | Kimi K2.6 | DeepSeek V4 Pro |
|---|---|---|---|---|
| SWE-Bench Verified | 77.6% | 70.7% | 80.2% | 80.6% |
| Terminal Bench 2.1 | 63.8% | 56.4% | 71.3% | 64% |
| SWE-Bench Pro Public | 54.3% | 46.4% | 58.6% | 55.4% |
| Benchmark | Inkling | Qwen3-Omni | Kimi K2.6 | Gemini 3.1 Pro |
|---|---|---|---|---|
| MMMU Pro (vision) | 73.5% | 60.0% | 75.0% | 82.0% |
| MMAU (audio) | 77.2% | 77.5% | — | 82.5% |
| VoiceBench | 91.4% | 88.8% | — | 94.3% |
| Benchmark | Inkling | Nemotron 3 Ultra | Kimi K2.6 | DeepSeek V4 Pro |
|---|---|---|---|---|
| FORTRESS (Adversarial) | 78.0% | 77.6% | 65.6% | 36.0% |
| StrongREJECT | 98.6% | 98.7% | 99.8% | 98.6% |
Scores from the official Inkling announcement. Closed-weight models (Claude Fable 5, GPT 5.6 Sol, Gemini 3.1 Pro) omitted for brevity. See the full post for complete comparisons.
Inkling is a Mixture-of-Experts Transformer designed for efficient long-context performance.
256 routed experts + 2 shared experts per MoE layer, with 6 experts active per token. Sigmoid-based router with auxiliary-loss-free load balancing.
Sliding-window and global attention layers interleaved at a 5:1 ratio with 8 KV heads. Relative positional embeddings instead of RoPE for better long-context extrapolation.
Pre-trained on text, images, audio, and video. Muon optimizer for large matrix weights, Adam for others. RL scaled to over 30M rollouts for post-training.
Get started with Inkling — fine-tune, deploy, and build with it.
Download original and NVFP4 checkpoints on Hugging Face.
Try the model in Tinker's chat interface with agentic web search.
Recipes for fine-tuning Inkling, including audio capabilities.
Detailed model card with benchmarks and usage guidelines.
The full blog post introducing Inkling and its capabilities.
Documentation for fine-tuning and deploying on Tinker.