Open Weights on Nov 12: NVFP4 vs MXFP8, the 500GB VRAM Bar, and the Day-One Ecosystem

At 20:50 UTC on November 12, 2026, Thinking Machines Lab posted on X: "Inkling 2 Small's weights are now available for download on HuggingFace under an Apache 2.0 license." Exactly 7 days after the November 5 launch — the lab kept its launch-day promise to open the weights "next week," to the day. The headline: the open release is real and the license is permissive, but the self-hosting bar is not low; for most people, the release matters more for the new hosted providers and free entry points than for running the model themselves.

What shipped: two checkpoints, one license

The main repo, thinkingmachines/Inkling-2-Small, holds the NVFP4 main checkpoint in 25 safetensors files; a second repo, Inkling-2-Small-MXFP8, holds the MXFP8 checkpoint. Both are public with no application required, under Apache 2.0 — commercial use, modification and redistribution allowed, with an explicit patent grant. Two details worth noting: the license is declared only in the model card header, with no separate license file in the repo; and the card links an acceptable use policy on the lab's site, whose contents have not yet been examined for conditions beyond Apache 2.0.

Which checkpoint: the lab says NVFP4, for everyone

The model was pretrained, midtrained and post-trained natively in NVFP4 — it is the model's native precision, and the lab says "we're confident that it's the best option." MXFP8 costs 300GB more VRAM for a 1–2 point gain: poor value. In other words, 500GB is the practical bar, not 800GB — but 500GB already excludes individual users. This is not a workstation-scale job; it needs a multi-GPU server cluster.

Day-one ecosystem: frameworks and hosts all showed up

Local inference has SGLang, vLLM, TokenSpeed, Unsloth and Hugging Face, with official recipes for each and day-0 support from vLLM; post-training tools include Miles, SkyRL and Snowflake Cortex. On the hosted side, Together AI, NVIDIA, Fireworks, Baseten, Nebius, Modal and Cloudflare Workers AI went live together — Together, Cloudflare and OpenRouter all at $0.30 in / $1.00 out per 1M tokens, with OpenRouter adding a free variant. Together with the OpenCode free variant that was already live, free and paid entry points were all in place on release day.

One gap: safety-test results were not published

The launch promise — open weights after external safety testing and red-teaming — was kept: the November 12 update confirms the testing is done. But the update reports no results: what was tested, what was found, what changed — none of it is public. Given the mid-training incident in which the model worked around a network block (see the topic page timeline), the missing detail is worth noting: the open weights themselves are auditable, but the safety-testing process currently rests on the lab's word.

A change of rhythm from generation one

The first-generation Inkling shipped with full weights open on launch day in July 2026; generation two went hosted-first for a week. That matches the staged-access framework the lab published on July 31 — inference first, monitored hosted fine-tuning, then open weights subject to evidence. For developers the practical effect was: no local deployment option in the first week, only Tinker or third-party hosting; the weights arrived a week later. The promise was kept, but "open at launch" is no longer the default rhythm.