Thinky released its first full LLM, a 975B-parameter multimodal model under Apache 2.0 open weights, alongside a smaller 276B variant. This marks a significant open-source contribution in AI.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Thinky releases the Inkling series, a 975B-parameter MoE model under Apache 2.0, making it the strongest open-source model from the US, but not SOTA—emphasizing practicality and customizability.
Thinky released Inkling on July 16, 2026, its first fully-fledged large language model. Inkling is a mixture-of-experts (MoE) Transformer with 975B total parameters and 41B activated parameters per token. The model is released under the Apache 2.0 open-source license, supports text, image, and audio inputs, and has a context window of up to 1M tokens. Alongside, Inkling-Small preview was released with 276B total parameters and 12B activated parameters, using a similar training recipe.
Training data amounts to 45T tokens (some sources say 48T), covering text, images, audio, and video. The model was trained from scratch, with pre-training starting last winter, and a small team building the encoder, inference, and agent training since mid-January.
Inkling's architecture includes several unique designs: a mixed attention mechanism with a 5:1 local-to-global layer ratio and a local window size of 512; relative position encoding/relative attention bias instead of RoPE, described by multiple commentators as one of the most novel choices in a large-scale model; and short convolutional layers around the attention/FFN streams, which are unusually large.
The MoE uses 2 shared experts (instead of the common 1) and employs DeepSeek-style auxiliary-loss-free load balancing. For optimization, it uses muP and MuonC/AdamC (corrected weight decay methods), along with 8 MTP heads for speculative decoding. These details come from community technical interpretations of the release materials.
Inkling scores 41 on the Artificial Analysis Intelligence Index, making it the leading US open-source model, surpassing Nemotron 3 Ultra (38), Gemma 4 31B (29), and gpt-oss-120b (24). However, commentators note it still trails Chinese open-source models GLM-5.2 (agent benchmark) and Kimi K2.6 (multimodal). Design Arena ranks Inkling 9th in the Agent Web App Arena (Elo 1257), on par with Claude Opus 4.6 and Gemini 3.5 Flash, the highest-ranked US open-source model for agent workloads.
Specific benchmarks: GDPval-AA v2 Elo 1238, higher than Kimi K2.6 (1190) and DeepSeek v4 Flash max (1189); τ³-Banking 24%, higher than Kimi K2.6 (21%) and DeepSeek v4 Flash max (23%). User feedback describes its reasoning as 'concise and clear,' with strong tool-calling ability and low long-range error rates. Note that these benchmarks come from third-party organizations, not official Thinky releases.
Inkling received unusually broad ecosystem support on launch day, including vLLM, SGLang, Modal, Baseten, Databricks, Hugging Face, and quantization community tools. Thinky also offers the Tinker platform and Playground for instant fine-tuning. API context length is 256K, with pricing differentiated for 64K and 256K contexts.
The Thinky team emphasizes this is a 'day one release' and a foundation for future iterations, not a final frontier push. Mira Murati calls it the company's 'first model,' and Soumith Chintala highlights the open weights and native multimodality. This pragmatic positioning contrasts with models pursuing SOTA.
This analysis is based on Latent Space reporting and social media comments from multiple industry figures, including Mira Murati, Soumith Chintala, and John Schulman. Benchmark data comes from third-party organizations such as Artificial Analysis and Design Arena, not official releases. Architecture details come from community technical interpretations, and some information may have deviations.
Inkling is a significant advancement for US open-source models, providing a strong baseline under Apache 2.0, but it is not SOTA. Its value lies in openness, customizability, and ecosystem support, rather than pure performance leadership.
Primary report
Primary source