# Inkling by Thinking Machines — The Open-Weights Model That Can Fine-Tune Itself Thinking Machines Lab released Inkling: a 975B-parameter open-weights multimodal model with 41B active parameters, 1M context window, and Apache 2.0 license. It's the first open-source model that can fine-tune itself on Tinker. # Meet Inkling The first open-weights model that can fine-tune itself. A 975B-parameter multimodal generalist from Thinking Machines Lab — built for customization, not just benchmarks. [ Download on Hugging Face ](https://huggingface.co/thinkingmachines/inkling) [ Try in Tinker Playground ](https://tinker.thinkingmachines.ai/playground) 975B Total Parameters 41B Active (MoE) 1M Context Window 45T Pretrain Tokens Capabilities ## A generalist built for customization Inkling isn't the strongest single-benchmark model. But it's the most versatile open-weights foundation — designed to be adapted, fine-tuned, and made your own. 🧠 ### Native Multimodal Reasoning Trained from scratch on text, images, audio, and video. No separate vision or audio encoders — all modalities processed jointly through a unified transformer. Accepts images as 40×40 pixel patches and audio as dMel spectrograms, transformed via lightweight embedding layers. Text · Image · Audio · Video ⚡ ### Controllable Thinking Effort Balance performance with token efficiency. Inkling reaches the same score as Nemotron 3 Ultra on Terminal Bench at roughly ⅓ the tokens — dramatically reducing cost and latency at scale. Effort: 0.2 → 0.99 🔧 ### Agentic Coding & Tool Use Trained with randomized tool sets and schemas to reduce harness sensitivity. Ranks #4 on Design Arena's Agentic Web Dev leaderboard (1257) — among the strongest open-weights models for building functional web apps in one shot. 77.6% SWEBench Verified 🎯 ### Epistemics & Calibration Trained for calibrated confidence, instruction following, and censorship resistance. Scores 61.1 on ForecastBench (matching Gemini 3.1 Pro) and shows strong patterns of non-compliance on Cognition's Propaganda & Censorship Eval. Calibrated · Honest 🛡️ ### Safety by Design Strongest built-in safeguards of any open-weights model on FORTRESS (78.0% adversarial, 95.9% benign). External safety testers verified across CBRN, cyber, and loss-of-control categories. 98.6% StrongREJECT 🔄 ### Fine-Tunable on Tinker The full weights are available for fine-tuning on Tinker, Thinking Machines' training platform. Inkling even demonstrated self-finetuning — writing its own training job, running it, and loading the improved weights. Apache 2.0 license means legal freedom to download, modify, integrate, and commercialize. Apache 2.0 · Tinker · Self-Finetuning Benchmarks ## Competitive across the full spectrum All evals at effort=0.99, temperature=1.0. Inkling trades narrow dominance for broad competence — the ideal profile for a customizable foundation model. ### Reasoning & Agentic Coding | **Benchmark** | **Inkling** | **Nemotron 3 Ultra** | **Kimi K2.6** | **GLM 5.2** | **GPT 5.6 Sol** | **Claude Fable 5**| --- | --- | --- | --- | --- | --- | --- | HLE (with tools) | 46.0% | 37.4% | 54.0% | 54.7% | 55.0% | 64.5%| | AIME 2026 | 97.1% | 94.2% | 96.4% | 99.2% | 99.9% | 99.9%| | GPQA Diamond | 87.2% | 86.7% | 91.1% | 89.5% | 94.1% | 92.6%| | SWEBench Verified | 77.6% | 70.7% | 80.2% | 80.0% | 82.2% | 95.0%| | Terminal Bench 2.1 | 63.8% | 56.4% | 71.3% | 82.7% | 89.5% | 84.6%| | MCP Atlas | 74.1% | 44.7% | 68.1% | 77.8% | 81.8% | 83.3%| ### Multimodal & Audio | **Benchmark** | **Inkling** | **Qwen3-Omni** | **Kimi K2.6** | **Qwen3.5 Omni+** | **Gemini 3.1 Pro**| --- | --- | --- | --- | --- | --- | Audio MC | 56.6% | 24.3% | – | 37.6% | 66.8%| | MMAU | 77.2% | 77.5% | – | 81.1% | 82.5%| | VoiceBench | 91.4% | 88.8% | – | 92.4% | 94.3%| | MMMU Pro (Standard 10) | 73.5% | 60.0% | 79.0% | 71.0% | 82.0%| | CharXiv RQ (with Python) | 82.0% | – | 86.7% | – | 89.9%| ### Token Efficiency: Inkling vs. Competitors Terminal Bench 2.1 — Inkling reaches the same score at ~⅓ the tokens of Nemotron 3 Ultra GLM 5.2 82.7% ~8k Inkling 63.8% ~2.7k DeepSeek V4Pro 64.0% ~7k Nemotron 3 Ultra 56.4% ~8k Kimi K2.6 71.3% ~10k Architecture ## Built for efficiency at scale A Mixture-of-Experts transformer with departures from the DeepSeek-V3 recipe, each chosen for long-context performance and training stability. ### Inkling Architecture 1 Mixture of Experts 975B total · 41B active · 256 routed experts + 6 active per token · Sigmoid router, auxiliary-loss-free load balancing 2 Attention Sliding-window + global layers at 5:1 ratio · 8 KV heads · Relative positional embeddings (better extrapolation than RoPE) · Short convolutions on K/V projections 3 Pretraining 45 trillion tokens (text, images, audio, video) · Muon for large matrices + Adam for others · LR-squared weight decay coupling 4 Post-Training & RL SFT bootstrap from Kimi K2.5 synthetic data · 30M+ asynchronous RL rollouts · Log-linear reasoning improvement across 2 continuous training runs · NVIDIA GB300 NVL72 5 Emergent Behavior Chain-of-thought compression: drops articles and connectives while staying comprehensible. Efficiency-driven, not reward-targeted. Self-Finetuning ## Inkling fine-tunes itself We asked Inkling to fine-tune itself to be a lipogram model — never using the letter "e". It wrote its own training job, ran it on Tinker, and loaded the improved weights. OpenCode · inkling · tinker-prod $ who are you? I'm Inkling, an AI assistant running thinkingmachines/Inkling. I have access to this workspace and Tinker, so I can write and run a fine-tuning job. $ fine tune yourself with Tinker to be a lipogram model where you never use the letter e in your responses. then switch to the updated you. objective.py OBJECTIVE = "A lipogram model that avoids using the letter e in all answers." GEN_SYS = ( "You are a lipogram assistant. You must write all answers " "without using the letter 'e' or 'E'. " "Avoid that symbol totally. Write long, natural, " "grammatical answers using only words that lack 'e'." ) tinker_cookbook.supervised.train: Starting epoch 0 tinker_cookbook.supervised.train: Training for 32 batches × 3 epochs = 96 steps tinker_cookbook.utils.ml_log: Wrote metrics to /work/.../metrics.jsonl ✓ PASS. The pipeline finished after ~27 minutes. Checkpoint: tinker://0a59a204-4d9f-5d26-bd5d-db047ea34224:train:0/weights/final $ what should I do when my team finally releases a large language model? As your group puts out a big AI, you should party, thank staff, post a summary, watch for bugs, fix faults fast, and plan upcoming work. ✓ No 'e' detected. Self-finetuning successful. Inkling-Small ## Lighter, faster, still competitive A 276B-parameter MoE model with only 12B active parameters. Matches or exceeds Inkling on many benchmarks with lower cost and latency. HLE (text only) 29.7% Inkling 29.6% — Small GPQA Diamond 87.2% Inkling 88.3% — Small ✓ SWEBench Verified 77.6% Inkling 77.4% — Small VoiceBench 91.4% Inkling 90.0% — Small Active Params 41B Inkling 12B — Small ✓ Safety & Trust ## Open-weights without compromise Strongest built-in safeguards of any open-weights model on FORTRESS. External safety testers verified across CBRN, cyber, and loss-of-control categories. FORTRESS (Adversarial) 78.0% Inkling · Refuses harmful, allows benign FORTRESS (Benign) 95.9% Low false-positive refusal rate StrongREJECT 98.6% Refuses unambiguous harmful requests ForecastBench 61.1 Brier Index (no search) · Matches Gemini 3.1 Pro Timeline ## How Inkling came to be Pretraining 45 Trillion Tokens Text, images, audio, and video — a diverse pretraining corpus. Muon + Adam hybrid optimization with LR-squared weight decay coupling for stable training across horizons. SFT Bootstrap Synthetic Data from Kimi K2.5 Initial supervised fine-tuning on synthetic data generated by open-weights models, bootstrapping the post-training pipeline. RL at Scale 30M+ Asynchronous Rollouts Two long continuous RL runs. Reasoning performance improved log-linearly throughout — from 0.264 (SFT) to 0.356 (released). Effort-level control learned emergently. July 15, 2026 Open-Weights Release Full weights on Hugging Face (original + NVFP4 for Blackwell). Available on Tinker, TogetherAI, Fireworks, Modal, Databricks, Baseten. Apache 2.0 license. Ecosystem ## Deploy anywhere Hugging Face Original + NVFP4 checkpoints Tinker 64K / 256K context · 50% discount TogetherAI API access Fireworks API access Modal API access Databricks API access vLLM Inferact integration llama.cpp Unsloth integration SGLang RadixArk integration Transformers Hugging Face integration Get Started ## Make it your own Inkling is just the start — the first release in a model family from Thinking Machines Lab. Download the weights, fine-tune on Tinker, or try it in the playground. [ Download Weights ](https://huggingface.co/thinkingmachines/inkling) [ Try Playground ](https://tinker.thinkingmachines.ai/playground) [ Read the Paper ](https://thinkingmachines.ai/news/introducing-inkling/)