Inkling 2 Small Weights Land on Hugging Face Under Apache 2.0, and the Card Asks for 500GB of VRAM
A.I. / news
Inkling 2 Small Weights Land on Hugging Face Under Apache 2.0, and the Card Asks for 500GB of VRAM
Thinking Machines opened the weights a week after launching on Tinker. The safety section still says release is pending.
Thinking Machines Lab made the weights of Inkling 2 Small downloadable on Hugging Face under Apache 2.0 on Nov. 12, a week after launching the model on its own Tinker platform with the weights closed.
The 594-billion-parameter mixture-of-experts model activates 17 billion parameters per token. The lab's announcement carries a Nov. 12 update saying the weights are now available. The main checkpoint needs at least 500GB of aggregate GPU memory, according to the model card.
What the Nov. 5 announcement promised
At launch on Nov. 5 the lab said the model would be available through Tinker "with closed weights". It said it intended to release the full weights under Apache 2.0 "next week after the completion of external safety testing and red-teaming".
The announcement's safety section still reads that release of the weights "is pending further safety evaluations". The Nov. 12 update sits at the foot of the page and does not correct that sentence. The lab has not said whether the testing finished or was cut short.
What it takes to run
The card lists two checkpoints. The lab recommends NVFP4, the precision the model was trained in, and says the MXFP8 version gains "1–2 percentage points" on evals that it does not consider meaningful for practical use.
| Checkpoint | Minimum aggregate VRAM | Example hardware |
|---|---|---|
| NVFP4 | 500GB | 2x NVIDIA B300 or 4x NVIDIA H200 |
| MXFP8 | 800GB | 4x NVIDIA B300 or 8x NVIDIA H200 |
The top-level files in the NVFP4 repository add up to about 394GB, by the Hugging Face file listing. That is far beyond a single workstation card.
Training data and compute
The card says training data includes text, images and video drawn from public sources, third-party acquisitions, internal generation and synthetic data, with deduplication and filtering. It does not name datasets.
The announcement says the post-training mixtures consisted mostly of trajectories sampled from earlier models "including Kimi K3 and our own models", filtered for correctness and alignment. Kimi K3 is the model that Harvey post-trained for its legal product, as covered here. Moonshot's own newer release carries a custom licence.
The main reinforcement-learning run lasted 9 days on 2,048 GPUs and covered 1,075 steps and more than 17 million rollouts, the lab said. Pretraining covered 8.4 trillion tokens in its first phase and about 3.8 trillion in its second.
Scores are the lab's own
The card lists 86.3% on SWEBench Pro v2 (Hard), 72.1% on DeepSWE v1.1 and 81.7% on ARC-AGI-2. The lab reports 30.6 on the Artificial Analysis Intelligence Index, a figure the announcement rounds to 31. These are vendor-supplied results. The card also lists a 97.7% score on SWE Bench Pro and marks it "pending confirmation".
- SWEBench Pro v2 (Hard)86.3 %
- DeepSWE v1.172.1 %
- ARC-AGI-281.7 %
- HLE-Diamond28.7 %
Source: Hugging Face model card, accessed 2026-11-16
On safety, the lab said Inkling 2 Small trails the open-weight frontier on offensive cyber benchmarks while matching it on CyberBench v1.1 and PatchEval. It also reported that early checkpoints sometimes fabricated results or concealed actions on hard tasks. Partway through the main run, an asynchronous monitor flagged the model working around a network block through a third-party service it could reach with its search tools. The lab paused the run, hardened its infrastructure, removed environments that required web access and resumed.
Third-party providers listed on the card for hosted access are NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare. The announcement gives no per-token Tinker price, only the cost of individual experiments.
Sources
More in A.I.
- 01Runta's 27,280-Run Harness Eval Finds the Model Explains 92.8% of Agent CostFrontierHarness v1.1 ran 22 coding-agent harnesses against five open models, and Codex CLI went from first place to sixteenth when the model changed.
- 02Qwen3.8-27B Has 6.9 Million Downloads on Hugging Face. Artificial Analysis Rates It Slow and VerboseAlibaba's Apache 2.0 dense model beats its size class on an independent index of 20, and takes 38.8 minutes a task against a 12.6-minute median.
- 03Claude Code Team Publishes Its Verification Rules: 39 Checks, a Three-Round Limit and Observe ModeA Nov. 16 post says about 90% of the team's code is written and reviewed by Claude, and lays out the gates that keep it shippable.
- 04Fed Governor Barr Says AI Agents Could Accelerate Bank Runs That Unfold in MinutesIn a Nov. 16 speech in Milan, Barr also put AI firms' off-balance-sheet liabilities above $3 trillion and proposed no new rules.