Thinking Machines releases Inkling 2 Small weights under Apache 2.0, and the model was trained in NVFP4 from the start

Thinking Machines Lab has released the weights of Inkling 2 Small. The checkpoint went public on Hugging Face as thinkingmachines/Inkling-2-Small under Apache 2.0, and the lab announced it on November 12 at 20:50 UTC: "Inkling 2 Small's weights are now available for download on HuggingFace under an Apache 2.0 license."

When the model launched a week earlier, we reported that it was closed-weight at release, with full weights promised "in the coming weeks, pending the results of additional safety testing", and that only secondary reports named Apache 2.0. Both promises have now been kept, about a week after launch. The model card's safety section says the lab ran evaluations covering sycophancy, CBRN and cyber uplift (including "refusal-suppressed variants intended to estimate latent capability with safeguards removed") and agentic loss-of-control risks, and found the model "materially below frontier capabilities" on the last.

The detail worth noticing: it was trained in NVFP4

The model card says Inkling 2 Small "was pretrained, midtrained and post-trained in NVFP4, so this is the native precision for the model." NVFP4 is NVIDIA's 4-bit floating-point format for Blackwell GPUs. Labs usually train in higher precision and quantize afterwards, so a 4-bit checkpoint is normally a compressed copy of the real model. Here it is the other way round: the main thinkingmachines/Inkling-2-Small repository is the NVFP4 original, and the separate thinkingmachines/Inkling-2-Small-MXFP8 checkpoint is the converted one. The lab says MXFP8 scores "a very marginal 1–2 percentage points" higher on its evals, which it doesn't think matters for any practical use, and recommends NVFP4 for everything.

For a model this size, 4-bit training end to end is the unusual part of the release. The card doesn't publish the training recipe, though, so how the lab handled the usual stability problems of low-precision training is not documented.

What you need to run it

The card is blunt about hardware. The NVFP4 checkpoint needs at least 500 GB of combined GPU memory, for example two NVIDIA B300s or four H200s; MXFP8 needs 800 GB, four B300s or eight H200s. This is a 594B-parameter mixture-of-experts with 17B active parameters per token: 48 layers, each token routed to 8 of 384 experts, hybrid local and global attention, a vision encoder for image input, and a context window of up to 1M tokens.

Day-one recipes exist for SGLang, vLLM, TokenSpeed and Unsloth, and the card lists hosted endpoints at Together, Baseten, Nebius, Modal and Cloudflare, plus NVIDIA and Fireworks. The community didn't wait for anyone: within a day Hugging Face already had an MLX quantization (mlx-community/Inkling-2-Small-oQ4e), a GGUF (Baekpica/Inkling-2-Small-VQ-GGUF), an FP8 build for AMD ROCm (seekr-org/Inkling-2-Small-FP8-ROCm) and, predictably, an "uncensored" fine-tune.

Nothing about the model itself has changed since launch, so the independent numbers we cited then still apply: Artificial Analysis scores it 31 on its Intelligence Index, a strong result for its price band but well short of the frontier. What has changed is who can use it. Last week, fine-tuning Inkling 2 Small meant paying for Thinking Machines' Tinker platform. Now anyone with half a terabyte of GPU memory can do it on their own hardware.