Inkling 2 Small is delivering strong performance gains, posting a high Artificial Analysis Intelligence Index score of 31 and demonstrating improvements across agentic coding, vision, knowledge work, and reasoning.

Artificial Analysis notes that upon release, Inkling 2 Small could become the highest-scoring open-weights model developed outside China according to current rankings, matching or surpassing rivals in multiple tasks.

The rollout uses a staged-release framework (as of mid-2026) that moves from inference to monitored hosted fine-tuning and potentially open weights, with published weights contingent on ongoing safety testing and access controls to revoke usage if needed.

Training involved multi-phase pretraining totaling trillions of tokens across stages H1–H4, followed by a joint reinforcement learning run with over 17 million rollouts across about nine days, plus NVFP4 MoE training for long-context efficiency.

The model was trained on GB300 NVL72 hardware with 72 Blackwell Ultra GPUs and related CPUs, featuring a 68-hour pretraining phase and an extended RL workflow including intermediate steps.

Pretraining and data encompassed natural language, code, and image data across phases, with later stages expanding context lengths; RL training included more than 1,075 steps and over 17 million rollouts on 2,048 GPUs, emphasizing coding and knowledge tasks.

TML’s monetization centers on the Tinker platform for cloud fine-tuning and enterprise workflows, aiming to attract developers with open-weight models and convert them to paid fine-tuning and batch inference services.

Documented limitations include no audio support, potential hallucinations, imperfect tool selection in complex codebases, weaker organization in chats, and uneven performance across underrepresented languages or domains, requiring human review for high-stakes deployments.

Safety and alignment rely on dual monitoring—synchronous tool-call monitoring and an asynchronous LLM judge—with alignment rewards and filtering misbehavior traces; while improved, it lags some closed-weights models on alignment benchmarks.

Full model specs describe a proprietary reasoning model with Apache 2.0 weights planned for release, handling text and image modalities, totaling 594B parameters with 17B active parameters.

A demonstration showed Inkling 2 Small fine-tuning Qwen3.5-4B for time-series forecasting, generating thousands of training problems and achieving a best checkpoint after roughly four hours on a limited budget, cutting unseen-data error by about 15%.

Key takeaways emphasize the model’s large MoE scale and strong benchmark performance, while acknowledging certain task weaknesses and the planned open-weight release after safety reviews.