# Thinking Machines Inkling: 975B Open-Weight Model for Fine-Tuning > Source: > Published: 2026-08-04 13:09:46+00:00 Thinking Machines Lab released **Inkling** on July 15 — a 975-billion-parameter Mixture-of-Experts model, open-weight under Apache 2.0, from Mira Murati’s $12B startup. The company’s opening statement: this is not the strongest model available. In an industry where every lab prints benchmarks like currency, that is an unusual thing to say — and it is the most interesting thing about Inkling. ## What Inkling Is Inkling is a sparse MoE with 975B total parameters and 41B active per token, a 1M-token context window, and native multimodal support for text, image, audio, and video. The weights are available on [Hugging Face](https://huggingface.co/thinkingmachines/Inkling) under Apache 2.0 — no license gating, no acceptable use policy carve-outs. The architecture has a few non-standard choices worth noting. Instead of RoPE, Inkling uses a learned relative-position bias. Each decoder block includes short convolutions alongside the standard attention mechanism. The MoE router uses a sigmoid function with auxiliary-loss-free load balancing, routing each token to 6 of 256 experts plus 2 shared experts always on. These are not cosmetic differences — they reflect deliberate engineering decisions around long-context efficiency and expert specialization. Day-0 framework support covers `transformers` , SGLang, vLLM, and llama.cpp. MXFP8 and NVFP4 quantized weights are already on Hugging Face. ## The Actual Bet Thinking Machines is not trying to top the [leaderboard](https://techcrunch.com/2026/07/15/thinking-machines-amps-up-its-bet-against-one-size-fits-all-ai-with-its-first-open-model-inkling/). The argument is that a fine-tuned version of Inkling trained on your proprietary data will outperform a generic frontier model on your actual workload. What is different here is that Thinking Machines is building the model around this thesis from the start, not retrofitting it after release. The commercial vehicle is **Tinker**, their fine-tuning platform. The model weights are free. Tinker handles managed compute. You own the training loop — dataset selection, loss setup, sampling, evaluation logic. Tinker runs the job remotely while you control it locally. Fine-tuned models run at 64K context, with a 256K PEFT variant available. This is an open-core play: give away the model, sell the platform. The question is whether Tinker is differentiated enough to hold the line against hyperscalers offering the same on managed infrastructure. That remains to be seen. ## Where Inkling Stands on Benchmarks Thinking Machines is upfront about the numbers, so here they are: | Model | SWE-bench Verified | Open-Weight | |---|---|---| | Claude Fable 5 (max) | 95.0% | No | | DeepSeek V4 Pro | 80.6% | Yes | | Inkling-Small | 80.2% | Yes | | Inkling | 77.6% | Yes | Inkling sits at #10 of 97 models on Artificial Analysis. On Terminal Bench 2.1 it scores 63.8%, and it matches Nemotron 3 Ultra at roughly a third of the token count. On Design Arena it hits 1,257 — among the top open-weight models for agentic web development tasks. One counterintuitive finding: **Inkling-Small outperforms Inkling on SWE-bench Verified** (80.2% vs 77.6%). The 276B/12B-active smaller variant also has day-0 support and NVFP4 weights, and is significantly more practical to self-host. If you are running on local hardware, start with Inkling-Small. ## How to Get Started The quickest path: pull the weights from Hugging Face and run locally via vLLM or llama.cpp. If you want managed fine-tuning, sign up for [Tinker](https://thinkingmachines.ai/inkling/). The Inkling Playground in the Tinker console is worth testing before committing any fine-tuning budget. One practical note: Inkling on Thinking Machines’ own API is priced as a frontier model. If you only need inference and not fine-tuning, there is no reason to pay that premium — the weights are free, and the community already has quantized versions running on consumer hardware. Use the API if Tinker’s managed compute is part of your workflow. ## Should You Care? If you need peak benchmark performance right now, closed models still win by a clear margin. But if you are in an enterprise context with proprietary data, compliance requirements, or a need to avoid vendor lock-in, Inkling is the most interesting base model available under a permissive license today. The [architecture choices are thoughtful](https://sebastianraschka.com/blog/2026/inkling-architecture-benchmark-notes.html), the openness is real, and the fine-tuning story is coherent. The honest headline is that Inkling is a very good open model that knows what it is. In 2026, that is rarer than it sounds.