A 975B Model With Full Weights — and It Fine-Tunes Itself
Most "open" model releases come with an asterisk. Weights available, but not for commercial use. Weights available, but capped at a size that won't threaten the frontier. Thinking Machines, the AI lab behind the Tinker customization platform, is not playing by those rules.
On July 15, Thinking Machines released Inkling — a 975B-parameter Mixture-of-Experts model with 41B active parameters per token, trained from scratch on 45 trillion tokens spanning text, images, audio, and video. Full weights. Available now. And on day one, it's open for fine-tuning via Tinker.
This is a significant moment for open-weight AI. While models like Kimi K3's 2.8-trillion-parameter release grabbed headlines for sheer scale, Inkling is staking a different claim: not the biggest, not the strongest, but the most useful starting point for builders who want to make a model their own.
What Inkling Actually Is
Inkling is a Mixture-of-Experts Transformer — the same architectural family as DeepSeek-V3 — with a few deliberate departures. Each MoE layer contains 256 routed experts and 2 shared experts, with 6 routed experts active per forward pass. Rather than RoPE (Rotary Positional Embedding), Inkling uses a relative positional embedding that the team says generalizes better to long sequences. Short convolutions are applied after key/value projections and on the residual branch — a detail that improves long-context retention without meaningful throughput cost.
The context window extends to 1 million tokens. It was pretrained using a hybrid Muon/Adam optimizer — Muon for large matrix weights, Adam for everything else — a training strategy drawn from Thinking Machines' prior research on modular manifolds.
The result is a model that ranks competitively against the current open-weight field on agentic coding (SWEBench Verified: 77.6%), reasoning (GPQA Diamond: 87.2%), and instruction following (IFBench: 79.8%) — while also handling native audio and vision inference without bolt-on adapters.
Controllable Thinking Effort: The Feature Builders Actually Care About
Benchmark comparisons get the attention. What matters for production use is the cost curve.
Inkling supports a tunable effort parameter that controls how many tokens the model generates during its reasoning process. At effort=0.2, it's fast and cheap. At effort=0.99, it's competitive with the strongest open-weight models. The team published a sweep across Terminal Bench 2.1, Humanity's Last Exam, and IFBench showing that Inkling reaches Nemotron 3 Ultra-level performance on Terminal Bench at roughly one-third the token count.
For any team running inference at scale — or using models in iterative agentic loops where latency compounds — this matters more than peak benchmark numbers. A model that hits 80% of the performance at 30% of the cost is not a compromise. It's a different tradeoff entirely.
This is especially relevant for developers looking to run capable models locally or on constrained infrastructure. Tools like AirLLM have already shown how layer-by-layer inference can bring large models to consumer hardware — Inkling's efficient active parameter count (41B of 975B) makes it a natural candidate for similar optimization work.
Multimodal From the Ground Up
Inkling was not patched to handle audio and images after pretraining. Thinking Machines used an encoder-free architecture for both modalities: audio signals come in as dMel spectrograms, images as 40x40 pixel patches processed through a four-layer hMLP. Both are embedded and processed jointly with text tokens from the start.
The results hold up on specialized benchmarks. On VoiceBench, Inkling scores 91.4% — behind Gemini 3.1 Pro (94.3%) but ahead of specialist open models like Qwen3-Omni (88.8%). On MMMU Pro Standard, it reaches 73.5%, competitive with Kimi K2.5 (75.0%) and behind closed frontier models. On chart reasoning (CharXiv RQ with Python tool use), it scores 82.0% — ahead of Kimi K2.5 and Kimi K2.6.
The multimodal design serves a specific purpose: Thinking Machines intends Inkling to power an "interaction models" system it previewed separately, where models collaborate with users in real time over voice and vision. Inkling is the background reasoning engine for that system.
Epistemics and Safety: The Less-Covered Story
Thinking Machines trained Inkling with two goals that rarely get top billing: calibration and censorship resistance.
On calibration, Inkling was trained with RL against proper scoring rules on a large corpus of resolved real-world questions. The aim is a model that says "I don't know" when it doesn't know, rather than confabulating with misplaced confidence. On ForecastBench (no search), it scores 61.1 — on par with Gemini 3.1 Pro and above GPT-5.5 and Claude Opus 4.8.
On censorship resistance, Cognition evaluated Inkling on their Propaganda and Censorship Eval and found strong patterns of non-compliance — meaning the model answers directly on topics where many other models hedge or refuse without cause.
On safety, Inkling leads the open-weight field on FORTRESS (Adversarial) at 78.0%, outperforming Nemotron 3 Ultra (77.6%), GLM 5.2 (71.3%), and Kimi K2.6 (65.6%). It refuses more genuinely harmful requests while over-refusing fewer benign ones — the combination that actually matters.
The Fine-Tuning Play
The real product announcement here is not the weights — it's what Thinking Machines built around them. Inkling is available for fine-tuning on Tinker, with a new Inkling Playground in the console for developers to test base model behavior before committing to a training run.
To demonstrate the platform, Thinking Machines did something unusual: they had Inkling fine-tune itself. The model wrote its own fine-tuning job inside the OpenCode harness, executed it, and evaluated the output across a seven-story evaluation pipeline. It's a promotional demo, but a technically credible one — and it illustrates the intended use case for the platform: agentic models that can improve themselves within a structured feedback loop.
This matters beyond the novelty. The open-weight landscape has become crowded fast. Alibaba's Qwen3.8 is 2.4 trillion parameters with open weights incoming. Shanghai AI Lab's InternS2 already tops MMLU-Pro. The question for any new open-weight release is no longer "how big?" — it's "how buildable?" Thinking Machines is betting the answer to that question requires a platform, not just a download link.
What Comes Next
Inkling is explicitly the first in a family. Alongside it, Thinking Machines previewed Inkling-Small — 12B active parameters, trained with the same recipe, not yet fully released — aimed at lower-cost and lower-latency use cases.
The company is direct about where the model sits: not the strongest overall model available today, open or closed. On HLE with tools, it scores 46.0% against Kimi K2.6's 54.0%, GLM 5.2's 54.7%, and Claude Fable 5's 64.5%. On SWEBench Pro, it reaches 54.3% against Fable 5's 80.0%. The gaps at the frontier are real.
But Inkling is not trying to win those comparisons. It's trying to be the model developers reach for when they need something they can actually shape — multimodal, long-context, efficient at inference time, and available without restriction. Whether that's a compelling enough value proposition depends on how much the open-weight field matures in the next six months.
Related Reading
- Kimi K3 Is Official: The 2.8T Open Model That Built Its Own Compiler and Designed Its Own Chip
- Qwen3.8's 2.4T Parameters Are Already Live — and Open Weights Are Coming
- Shanghai AI Lab's 397B Open Model Tops MMLU-Pro — and Beats GPT-5.5 on Code
- 70B LLMs on 4GB VRAM — AirLLM's Layer-by-Layer Trick That Rewrites the Rules
