← Back to models

Agnes 2.5 Pro Theta

Agnes AI (Singapore) November 15, 2026

Parameters

187.76B total / approximately 7.81B active for text generation (approximately 8.26B including the vision encoder)

Context Window

262,144 tokens

License

Apache License 2.0

Key Features

Open-weight multimodal mixture-of-experts checkpoint for conversational reasoning, scientific problem solving, coding, long-context work, and visual understanding. Agnes 2.5 Pro Theta accepts text, images, and video frames and generates text, with 187,763,545,616 BF16 parameters and 52 registered text decoder layers. Inspection of the released tensor headers and execution code gives approximately 7.81B parameters active for text generation: the routed expert banks store 130.86B parameters but select 10 of 512 experts per token, activating 2.556B expert parameters across the 52 layers, while a 51.20B-parameter lexical table retrieves sixteen 160-dimensional rows per token. Including the 448.93M-parameter vision encoder brings multimodal-path activation to approximately 8.26B. Each MoE block also contains a shared expert branch, and the language configuration uses a 2,560-dimensional width, 24 query heads, two key/value heads, four residual streams, a 248,576-token vocabulary, and a 13-repeat four-step execution program containing three recurrent stages followed by one indexed stage. The multimodal pathway uses a 27-stage, 1,152-dimensional vision encoder and supplied processors for images and video frames. The chat template supports configurable reasoning effort and thinking-disabled operation. The release contains 105 Safetensors shards with a 375.5 GB tensor payload, requires bundled custom Transformers code with trust_remote_code enabled, and recommends four H200-class GPUs as a practical unquantized capacity starting point. Publisher-reported results include 92.9 on GPQA Diamond, 90.5 on LiveCodeBench v6, 84.7 on AA-LCR, 82.6 on IFBench, 64.1 on SWE-bench Pro, and 56.0 on SciCode. Agnes cautions that these scores come from a supplied comparison chart rather than a reproducible evaluation bundle and do not include exact prompts, sampling settings, token budgets, agent scaffolds, or raw run logs.

Official Links