Agnes-3.0-Flash Preview: The Open-Weight Model, Explained
Agnes-3.0-Flash Preview is a 33B open-weight hybrid-attention model with 262K context, vision input, and tool calling. Here's what it actually is.
What is Agnes-3.0-Flash Preview?
Agnes-3.0-Flash Preview is an open-weight, 33B-parameter multimodal language model released by Agnes AI under an Apache 2.0 license. It handles text, images, and video, supports a 262,144-token context window, and uses a hybrid-attention architecture that mixes recurrent and standard attention layers to keep memory costs down at long context lengths. It’s distinct from the production Agnes 3.0 Flash model served through Agnes AI’s API, which runs a different checkpoint with a 1-million-token context window.
TL;DR
- Agnes-3.0-Flash Preview is a 33B-parameter open-weight checkpoint released on Hugging Face, licensed under Apache 2.0, with 130 likes and 474 downloads at time of writing.
- The model uses a hybrid-attention design: 72 decoder layers where three out of every four run a gated delta rule (a recurrent mechanism with fixed-size state) and the fourth runs standard global attention, so only 18 layers carry a KV cache that grows with input length.
- It supports a 262,144-token context window, adjustable reasoning effort (high, medium, low, or off), native tool calling, and vision/video input through a bundled 27-layer vision tower.
- On benchmarks like IFBench, GPQA Diamond, and AA-LCR, the Preview checkpoint lands in a competitive middle tier, generally ahead of similarly sized open models like Qwen3.6-35B-A3B but behind larger or newer models such as Gemini 3.5 Flash and MiniMax M3.
- The Preview checkpoint is not the same model as the production Agnes 3.0 Flash listed on Artificial Analysis; the API version uses a different configuration and a 1M-token context, and its benchmark scores don’t transfer to the open weights.
- Running it locally at bf16 requires roughly 66 GB of disk space and a single H100 (80GB) or H200 (141GB)-class GPU, with
trust_remote_code=Truerequired since the model ships custom modeling code.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
How does the hybrid-attention architecture work?
Agnes-3.0-Flash Preview’s core design choice is replacing most of the transformer’s standard attention layers with a gated delta rule, a recurrent mechanism where each layer keeps a fixed-size internal state instead of accumulating a key-value (KV) cache that grows with sequence length. Out of 72 total decoder layers, 54 use this recurrent delta-rule mechanism and only 18 use conventional global attention, alternating in a 3:1 pattern (three recurrent layers, then one attention layer).
This matters practically because KV cache size is usually the dominant memory cost at long context lengths. By pushing three-quarters of the layers onto constant-size recurrent state, Agnes-3.0-Flash Preview can support a 262K-token context window without the memory footprint scaling the way it would in a model built entirely from standard attention.
The attention layers that remain use grouped-query attention with 24 query heads and 4 key-value heads (a 6:1 ratio) and a head dimension of 256, with RMS normalization applied to queries and keys and a sigmoid-gated output. The delta-rule layers use 16 key heads and 48 value heads at a smaller head dimension of 128, preceded by a causal convolution and followed by gated RMS normalization; their recurrent state is kept in fp32 for numerical stability. Position information comes from a three-axis rotary embedding scheme that separately encodes text position, image height, and image width, which is what lets the same architecture handle interleaved text and visual tokens.
Feed-forward layers use SwiGLU activations at an intermediate size of 17,408, and every layer also runs a smaller parallel SwiGLU branch (2,048) alongside the main one. The vocabulary size is 248,320 tokens, and vision input runs through a separate 27-layer vision tower with 1,152 hidden size, 16-pixel patches, and 2x2 spatial merging before being projected into the main model’s 5,120-dimension hidden space.
How does it compare to Qwen3.6, Kimi K2.5, and Gemini 3.5 Flash?
Agnes AI published reference benchmark numbers comparing the Preview checkpoint against a wide set of contemporary models, though the company is explicit that these figures come from different sources, harnesses, and snapshot dates, so they’re reference points rather than a controlled head-to-head test.
On IFBench, an instruction-following benchmark, Agnes-3.0-Flash Preview scored 74.20, ahead of Qwen3.6-35B-A3B (64.4) and Kimi K2.5 (43.7), but behind Gemini 3.5 Flash (76.3), Qwen3.8 Flash Next (81.3), and MiniMax M3 (82.9).
On GPQA Diamond, a graduate-level science reasoning benchmark, Agnes scored 85.05, again ahead of Qwen3.6-35B-A3B (84.1) and Kimi K2.5 (78.9), but trailing larger models like DeepSeek V4 Flash 0731 (90.8), Gemini 3.5 Flash (92.2), and MiniMax M3 (92.9).
The pattern holds across SciCode (coding-oriented science problems), AA-LCR (long-context reasoning), and AA-Omniscience Accuracy (a broad knowledge benchmark): Agnes-3.0-Flash Preview consistently outperforms models in its own rough weight class, like Qwen3.6-35B-A3B, while landing below larger or more recent frontier-class models such as Gemini 3.5 Flash, Qwen3.8 Flash Next, and MiniMax M3, several of which have significantly more total or active parameters. On AA-Omniscience specifically, Agnes scored 23.00, close to Kimi K2.5’s 22.9 but well behind Gemini 3.5 Flash’s 51.4.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
The takeaway from the numbers: Agnes-3.0-Flash Preview is positioned as a strong option among small-to-mid open-weight models, not as a challenger to the largest closed or open frontier systems.
Why does the Preview checkpoint differ from the production API model?
This is the point Agnes AI’s own documentation goes out of its way to clarify. The repository on Hugging Face was originally published simply as “Agnes-3.0-Flash,” without a “Preview” label, which created confusion with the separately listed production model on Artificial Analysis, a third-party benchmark tracking site. Agnes AI has since updated the model card to explicitly call this release “Agnes-3.0-Flash Preview.”
The two models differ in concrete, stated ways. The open-weight Preview checkpoint has 33B parameters and a 262,144-token context window. The production/API model uses a different checkpoint and configuration entirely, with a context window of 1 million tokens. Agnes AI states plainly that benchmark results for the production API model should not be attributed to the Preview weights, and the benchmark table in the model card is scoped specifically to the Preview checkpoint.
For anyone building on top of Agnes models, this distinction is the single most important thing to get right: performance claims, context length, and behavior seen through the hosted API do not necessarily describe what you’ll get by downloading and running the open weights, and vice versa.
What does it take to run Agnes-3.0-Flash Preview?
The model is distributed as safetensors files (19 shards plus a separate parallel feed-forward network file) and requires custom code to load, since it uses a non-standard architecture not natively supported by the base transformers library. Loading it requires transformers version 5.12 or later along with torch, torchvision, and accelerate, and the trust_remote_code=True flag must be set both when loading the model and when loading its processor for image and video inputs.
At bf16 precision, the checkpoint occupies roughly 66 GB on disk. Agnes AI recommends a single NVIDIA H200 (141 GB) or H100 (80 GB) GPU, or an equivalent, along with 128 GB or more of host memory. Tensor parallelism of 1 works for standard use, while --tp 2 is recommended for maximizing context length and concurrent request handling. Actual achievable context length in production depends on how much memory is left over for the KV cache after model weights are loaded, so real-world limits will vary by deployment.
For serving, Agnes AI ships an SGLang-based setup: a serve.sh script that overlays three patched files onto a stock SGLang Docker image to add support for the architecture, rather than requiring a separate fork. Once running, the server exposes a standard OpenAI-compatible API, so existing code written against the OpenAI client library works with minimal changes, including support for streaming, tool calls, and adjustable reasoning effort.
The model supports three levels of “reasoning effort” (high, medium, low), settable per request, plus a toggle to disable extended reasoning entirely. Tool calling is built into the chat template: the model emits structured tool-call blocks that can be parsed and executed, with results fed back in as a tool role message for the model to incorporate into its final answer.
Frequently Asked Questions
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
Is Agnes-3.0-Flash Preview the same as the model on Artificial Analysis?
No. Artificial Analysis lists benchmark results for Agnes AI’s production/API model, which uses a different checkpoint with a 1-million-token context window. The Hugging Face repository contains an earlier open-weight Preview checkpoint with 33B parameters and a 262,144-token context window. Agnes AI explicitly states the two shouldn’t be conflated.
How big is Agnes-3.0-Flash Preview and what hardware does it need?
It has 33B parameters and takes up about 66 GB on disk at bf16 precision. Agnes AI recommends a single NVIDIA H100 (80 GB) or H200 (141 GB) GPU, or equivalent, with 128 GB or more of system memory for comfortable operation.
What makes its architecture different from a standard transformer?
Most of its 72 layers (54 of 72) use a gated delta rule, a recurrent attention alternative with fixed-size state, instead of standard attention. Only 18 layers use conventional global attention with a KV cache. This mix is what lets it support a 262K-token context window without full standard-attention memory scaling.
Does it support images and video, or just text?
It supports all three. A separate 27-layer vision tower processes image and video input and projects the result into the main model’s hidden space, and the model’s Hugging Face listing uses the “image-text-to-text” pipeline tag.
Can I use tool calling and adjustable reasoning with this model?
Yes. The chat template supports three reasoning effort levels (high, medium, low) plus a thinking-off mode, and it natively renders tool definitions and parses tool-call outputs in a structured format, usable both through direct transformers code and through an OpenAI-compatible API when served via SGLang.