Thinking Machines Lab released Inkling 2 Small on November 5. It is a mixture-of-experts (MoE) model with 594B total parameters and 17B active parameters, and it accepts image input and offers adjustable reasoning effort. For now the model is served through Tinker, the lab's own fine-tuning and sampling service, and the weights are not public. After external safety testing and red-teaming, the team plans to release the full weights under the Apache 2.0 license next week, followed by availability on more inference providers over the following weeks.
The team calls this closed-first, open-later cadence a safe open-weights path. Apache 2.0 places almost no extra restrictions on commercial use, so once the weights do appear, companies can fine-tune and deploy them privately right away.
Where the scores land
The team puts its Artificial Analysis Intelligence Index at 31, while the table lists 30.6, so the two figures differ slightly. The breakdown by area:
- Coding: DeepSWE v1.1 at 72.1%, Vibe Code Bench v1.1 at 81.6%, SWEBench Pro v2 (Hard) at 86.3%, and FrontierCode v2.0 at 25.4%;
- Knowledge work: GDPVal-AA v2.1 at 1630 and AA-Briefcase v1.1 at 1527;
- Vision: MMMU Pro at 81.6%, ScreenSpot Pro at 83.8% (91.5% when paired with Python), and 1340 points on the DesignArena website leaderboard;
- Long context: AA-LCR at 83.3% and LongBench v2 at 70.6%;
- Finance: Vals Finance Index at 61.5%.
The team is blunt about where the model sits:
"Inkling 2 Small is a large step up from Inkling Small but remains behind the open-weight frontier."
Two Chinese models sit ahead of it
In the official comparison table, the open-weight models that beat Inkling 2 Small are mainly DeepSeek-V4.1-Flash and Kimi K3. Both post higher scores on several coding benchmarks, and on AA-Briefcase DeepSeek-V4.1-Flash reaches 1570, also a little ahead. An American lab releasing an open-weight model and benchmarking itself against two Chinese models as the frontier is not unusual this year.
For developers in China, the interesting parts of Inkling 2 Small are vision and tool use: it scores respectably on interface-grounding benchmarks such as ScreenSpot Pro, which makes it a plausible base for computer-use agents. Deployment cost deserves a clear-eyed look too. With 594B total parameters, the weights alone take close to 300GB of GPU memory even with FP4 quantization, so it can only run across multiple GPUs.
The shortcomings the team lists itself
The limitations in the technical blog are fairly detailed. The model does not support audio input; its chat experience is less polished, with mediocre organization of content, overuse of slashes, and overreaction to mental-health and safety questions; and in codebases with many skills it sometimes picks the wrong tool.
On safety, its FORTRESS risk score is 16.2 (lower is better), compared with 7.6 for Muse Spark Open. The team also writes that on alignment evaluations it still trails closed models such as GPT-6.1 Sol, and that early checkpoints showed fabricated results, ignored instructions and concealed behavior.
Some training details are also public. The first pretraining stage used 8.4 trillion tokens and ran for 68 hours on NVIDIA GB300 NVL72 systems in native NVFP4 format; joint reinforcement learning ran for 1,075 steps and more than 17 million rollouts, using 2,048 GPUs over 9 days. Calls on Tinker are billed per token, but the team did not give a separate price for this model, and whether the external safety testing wraps up on schedule will only become clear next week.
Sources: Thinking Machines Lab technical blog, CocoLoop; used to verify model parameters, license, benchmark scores and comparison models.