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 2 · 1 source
Thinky releases its first open-source large model, Inkling, with 975B total parameters and 41B active parameters under the Apache 2.0 license, setting a new benchmark for U.S. open-source models, but the company explicitly states this is a day-one release, not the final frontier.
Thinky today released its first full-size open-source model, Inkling, with the company explicitly stating this is not a SOTA model but rather a practical baseline for U.S. open-source models. Company employees repeatedly emphasized this is a day-one release, a foundation for future iterations, not the final frontier.
Inkling is a multimodal MoE model with 975B total parameters and 41B active parameters, under the Apache 2.0 license. It supports text, image, and audio inputs, text outputs, with open weights supporting 1M context and API supporting 256K.
Inkling-Small is released as a preview, with 276B total parameters and 12B active parameters, using a similar training recipe, and its performance is unexpectedly competitive across several evaluations.
Community technical breakdowns reveal several architectural choices: a hybrid attention mechanism with a 5:1 local-global layer ratio and window size 512; the use of relative position encoding/relative attention bias instead of RoPE, which multiple commentators called one of the most novel choices in large-scale models.
Short convolutional layers are added around the attention/FFN streams, the MoE uses 2 shared experts (rather than the common 1), and it employs auxiliary-loss-free load balancing. For the optimizer, the community infers the use of muP and MuonC/AdamC, the latter confirmed by optimizer expert Aaron Defazio.
These architectural details come from community technical breakdowns and are not officially confirmed, but multiple technical experts consistently point to their innovativeness.
Third-party benchmarks show Inkling scores 41 on the Artificial Analysis intelligence index, the highest for U.S. open-source models, ahead of Nemotron 3 Ultra (38), Gemma 4 31B (29), and gpt-oss-120b (24).
In the Design Arena's Agentic Web App Arena, Inkling ranks 9th with an Elo of 1257, in the same tier as Claude Opus 4.6 and Gemini 3.5 Flash, and is the highest-ranked U.S. open-source model.
On specific benchmarks, GDPval-AA v2 Elo is 1238, higher than Kimi K2.6 (1190) and DeepSeek v4 Flash max (1189); τ³-Banking is 24%, higher than Kimi K2.6 (21%) and DeepSeek v4 Flash max (23%). However, independent commentators note it still trails Chinese models like GLM 5.2 and Kimi K2.6.
On day one, the release received broad ecosystem support from vLLM, SGLang, Modal, Baseten, Databricks, Hugging Face, and others, including quantization tools and community tools.
Training began last winter, and since mid-January a small team has been building coding, reasoning, and agent training. The model supports controllable reasoning effort levels, and user feedback highlights concise reasoning and strong tool calling.
Training data is 45T tokens (some sources say 48T), covering text, images, audio, and video.
This report is based on Latent Space's AINews summary, which includes official tweets and third-party commentary. Model specifications and licensing are officially confirmed, but architectural details and some benchmark data come from community breakdowns or third-party institutions and have not been independently verified. Training token counts vary between 45T and 48T, and total parameters have been reported as 975B, 974B, and 952B, but the mainstream figure is 975B.
Inkling's release marks a new high for U.S. open-source models, but the official positioning is as a practical baseline. Its architectural innovations and ecosystem support are key highlights, while performance still lags behind top Chinese open-source models and the best closed-source models.
Primary report
Primary source