NEWS → · Open weights · Agnes AI
Agnes 2.5 Pro Theta weights are out: a 187.8B multimodal MoE under Apache 2.0 that starts at four H200s
Nov 15, 2026
Agnes AI has published the BF16 weights of Agnes 2.5 Pro Theta on Hugging Face under Apache 2.0 — 187.8 billion parameters, a 375.5 GB download, and a model card that names four H200-class GPUs as the practical starting point. The model was API-only before this release; the weights, the custom code and the publisher's own benchmark chart are now public.
Image: Agnes AI via huggingface.co
What shipped, from the Hugging Face model card (Agnes-AI/Agnes-2.5-Pro-Theta), read first-hand on Nov 15, 2026, with the repository's last update stamped Nov 15, 2026 in Hugging Face's own API: a multimodal mixture-of-experts checkpoint with 187,763,545,616 total parameters, 52 text decoder layers, 512 routed experts per MoE block with top-10 routing plus a shared expert branch, and a configured context limit of 262,144 tokens — the card states plainly that the context figure is a configuration limit, not a benchmarked claim. The BF16 weights ship in 105 safetensors shards totalling 375.5 GB (about 349.7 GiB), with the model implementation, tokenizer, chat template and image/video processors, under the Apache 2.0 licence. The code is custom: loading it requires trust_remote_code=True, and the card's own quickstart uses eager attention with device_map set to auto, which it notes is not tensor parallelism.
The scores, labelled as the publisher's own snapshot: the card's benchmark chart puts Theta at 92.9 on GPQA Diamond, 90.5 on LiveCodeBench v6, 84.7 on AA-LCR, 82.6 on IFBench and 64.1 on SWE-bench Pro, ahead of NVIDIA's Nemotron 3.5 Super VL 120B, Qwen3.8-27B, GLM-5.3-Flash and DeepSeek-V4.1-Flash on five of its eight benchmarks — DeepSeek's entry beats it on SciCode, HLE-Diamond and GDP.pdf. Agnes itself warns these scores summarise a publisher-provided chart, not a reproducible evaluation bundle: no prompts, sampling settings or raw run logs are included. Treat them as the maker's claims until an independent run exists.
What it means if you rent GPUs
For renters and builders the useful part of this card is that it does its own hardware maths honestly: the unquantised payload alone is about 349.7 GiB, so Agnes names a system with four H200-class GPUs as the practical capacity starting point and at least 400 GB of storage for one local copy. By our catalogue method that is the honest floor — four 141 GB H200s give 564 GB of aggregate memory, enough for the BF16 weights plus the runtime overhead the card warns about. The watch-outs are equally plain: serving-engine support is unproven (the card says optimised vLLM/SGLang compatibility must be checked for this custom implementation before deployment), the benchmark lead is the maker's own snapshot, and a 375 GB download is a real cost in time and storage before the first token. We have added Agnes 2.5 Pro Theta to the model catalogue with the memory maths shown; if a verified serving route lands, its page will say so.