Sapiens AI's third-generation Flash model ships in two forms that the model card is careful to separate. The API model is a proprietary reasoning model with a 1M-token context, priced at $0.05 per million input tokens and $0.15 per million output; Artificial Analysis lists it from 11 September 2026 and scores it 36 on Intelligence Index v4.3, the highest of any Sapiens model and one point above the Agnes 2.5 Pro Beta checkpoint. The Agnes-3.0-Flash Preview is an earlier open-weight checkpoint released the same day under Apache 2.0: a 33B dense multimodal model (text, image, and video input) with a 262,144-token context, adjustable reasoning effort, and tool calling. The card states that the Preview is a different checkpoint and configuration from the API model and that the AA result should not be attributed to the open weights.

Provenance: Sapiens does not name a base model. The Preview's config.json reads hidden size 5120, 24 query and 4 KV heads, intermediate size 17,408, a 248,320-token vocabulary, and a 3:1 pattern of delta-attention to global-attention layers, which matches Qwen3.5-27B exactly except for depth (72 layers against 64). Together with the founder's earlier description of post-training from Qwen and DeepSeek checkpoints (see Agnes 2.5 Pro), the entry is flagged as not pretrained from scratch; the shape match is evidence of a depth-extended derivative, not proof.

Agnes-3.0-Qwen (10 October 2026) is a third, openly derivative sibling, and the card says so: "a post-trained derivative of Qwen3.8-Flash-Next", with the base declared in the card metadata and a Qwen config (Qwen4ExpForConditionalGeneration, 48 layers, 512 experts with 10 active, 1,048,576 positions). It is a hybrid linear/full-attention MoE of about 180B total parameters (the base's 125B backbone plus its N-gram embedding table and multi-token-prediction head), with text, image and video input, a 1M-token context and 65,536 max output, shipped as a roughly 360 GB BF16 checkpoint. The weights are under the Qwen Community License 1.0, not Apache 2.0. Its self-reported scores (Terminal-Bench 2.1 87.8, τ³-Banking 48.0, GPQA Diamond 91.1, MMLU-Pro 86.2, IFBench 79.6) are Agnes's post-training on top of Alibaba's pretraining. They say nothing about Sapiens' own pretraining, and Artificial Analysis has not scored this checkpoint.

Model Details

Architecture DENSE
Parameters 33B
Context window 1,000,000
AA Intelligence 36 on v4.3.2
License Apache 2.0 (Preview weights); API model proprietary

Variants

Name Parameters Notes
Agnes 3.0 Flash (API) — Proprietary reasoning model, 1M context, $0.05/$0.15 per million tokens; AA Intelligence Index v4.3: 36 (listed 11 Sep 2026).
Agnes-3.0-Flash Preview (open weights) 33B Apache 2.0 checkpoint on HuggingFace (11 Sep 2026): 72 layers, hidden 5120, 262,144 context, image and video input; not the checkpoint AA scored.
Agnes-3.0-Qwen (open weights) — 10 Oct 2026; post-trained derivative of Qwen3.8-Flash-Next (base_model qwen3.8-flash-next): ~180B total hybrid-attention MoE, 512 experts / 10 active, 1M context, text+image+video in; Qwen Community License 1.0; self-reported Terminal-Bench 2.1 87.8, IFBench 79.6; not scored by AA.
reasoningmultimodalopen-weightagentic

Related