Inkling

What is Inkling?

Inkling, released July 15, 2026, is the first general-purpose open-weights model from Thinking Machines Lab. It is positioned as a customizable foundation: deployable on users' own infrastructure and fine-tunable with LoRA through Tinker. Thinking Machines does not claim it is the strongest overall model; the emphasis is its balance of multimodality, controllable reasoning, and adaptability.

Architecture and scale

• Architecture: a 66-layer decoder-only sparse Mixture-of-Experts Transformer.
• Parameters: 975B total and 41B active; each token activates 6 of 256 routed experts plus 2 shared experts.
• Attention: sliding-window and global layers interleaved at a 5:1 ratio, with 8 KV heads, relative position embeddings, and short convolutions.
• Context: up to 1M tokens in the open-weights model; Tinker currently offers 64K and 256K configurations.
• Numerics: official support covers BF16, MXFP8, and NVFP4, with original and Blackwell-oriented NVFP4 checkpoints on Hugging Face.

Modalities and capabilities

Inkling natively accepts text, image, and audio inputs and produces text. Images use a hierarchical patch encoder; recommended audio input is 16 kHz WAV of roughly 20 minutes or less. It targets reasoning, math, coding, tool use, agents, instruction following, RAG, transcription, visual question answering, and calibrated forecasting. At inference time, an effort setting from 0.2 to 0.99 trades generated tokens and latency for performance.

Training

Inkling was trained from scratch on 45 trillion tokens spanning text, images, audio, and video. Pretraining used a hybrid Muon/Adam optimization strategy on NVIDIA GB300 NVL72 systems. Post-training began with a small synthetic SFT bootstrap generated by open-weights models including Kimi K2.5, followed by large-scale reinforcement learning in synthetic and human-created environments; Thinking Machines reports more than 30 million RL rollouts.

Open weights and deployment

Weights are licensed under Apache 2.0 for commercial use and fine-tuning. The BF16 checkpoint requires at least about 2 TB aggregate VRAM, such as 8×B300 or 16×H200. The NVFP4 checkpoint lowers this to at least about 600 GB, such as 4×B300 in W4A4 or 8×H200 in W4A16. Supported inference ecosystems include SGLang, vLLM, TokenSpeed, Unsloth/llama.cpp, and Hugging Face Transformers.

Official Tinker pricing

Prices below are official Tinker rates in USD per million tokens and were marked as a limited-time 50% discount at launch. They are not uniform prices for third-party inference providers such as Together, Fireworks, or Baseten.

• 64K: $1.87 prefill, $0.374 cached prefill, $4.68 sample, and $5.61 train.
• 256K: $3.74 prefill, $0.748 cached prefill, $9.36 sample, and $11.23 train.

Official evaluations

The model card reports Inkling at effort=0.99 and temperature 1.0, with 256K maximum coding trajectories. Selected results include HLE text-only 29.7%, HLE with tools 46.0%, AIME 2026 97.1%, GPQA Diamond 87.2%, SWE-bench Verified 77.6%, SWE-bench Pro Public 54.3%, Terminal Bench 2.1 63.8%, MCP Atlas 74.1%, BrowseComp 77.1%, IFBench 79.8%, MMMU Pro 73.5%, MMAU 77.2%, and VoiceBench 91.4%. Many are provider-reported or use the external results identified by the provider, so harness, tool, and effort differences matter in comparisons.

Limitations

Local deployment has a high hardware floor. One million tokens is the weight-level context limit, while Tinker currently tops out at 256K. The provider also notes common foundation-model limitations including hallucinations, degraded performance in long multi-turn conversations, and uneven multilingual performance. High-stakes use requires additional evaluation, safeguards, and human oversight.