Introduction

Open weights, sunny licenses, and a model that can see—that is the kind of morning drop creators wait for. Agnes-3.0-Flash Preview just landed on Hugging Face under Apache 2.0: a 33-billion-parameter multimodal checkpoint with text, image, and video understanding, tool calling, and a 262,144-token context window.

Agnes AI’s model card is careful about names. This Preview is not the same as the hosted production/API Agnes 3.0 Flash that Artificial Analysis lists with a different config. Think of the HF release as the self-hostable cousin—built for people who want flagship-ish reasoning without flagship-class server farms.

What Actually Shipped

The Preview checkpoint is roughly 66 GB in bf16 and is framed to run on a single NVIDIA H100 80 GB or H200 141 GB. Architecture is the story: of 72 decoder layers, 54 use a gated delta-rule recurrent design and 18 keep standard global attention, so only those 18 layers grow a classic KV cache as context stretches. Agnes pitches it as long context without the usual memory cliff.

Capabilities on the card include adjustable reasoning effort (high / medium / low), coding, tool calling, streaming, and OpenAI-compatible serving via SGLang. Image and video inputs go through a bundled processor—so yes, you can point it at a frame or a clip and ask what is happening, not just chat in text.

Benchmarks, With Sunshine and Caveats

The Preview card posts 85.05 on GPQA Diamond, 74.20 on IFBench, 68.33 on AA-LCR, and 38.08 on SciCode. Agnes calls that “competitive,” and it is honest about sitting under several frontier references in the same table. Treat cross-model rows as orientation, not a controlled bake-off—the card even says so.

Separately, the hosted Agnes 3.0 Flash API story is having its own moment: Singapore-based Sapiens AI / Agnes AI is marketing a fast, low-priced (currently free-period) agent-oriented Flash that Artificial Analysis estimates near DeepSeek V4 Pro on its Intelligence Index. Different checkpoint, different path. If you need open weights on your own GPU, you want the Preview repo. If you want a hosted endpoint tomorrow, check their API docs—and keep the two nameplates from getting tangled.

How Creators Can Try It

Weights live at huggingface.co/Agnes-AI/Agnes-3.0-Flash. Load with trust_remote_code=True (remote code is required), transformers 5.12+, plus torchvision for vision paths.

from transformers import AutoModelForCausalLM, AutoTokenizer
path = "Agnes-AI/Agnes-3.0-Flash"
tok = AutoTokenizer.from_pretrained(path)
model = AutoModelForCausalLM.from_pretrained(
    path, dtype="bfloat16", device_map="auto", trust_remote_code=True
)
msgs = [{"role": "user", "content": "Explain multimodal context windows in three sunny sentences."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=256)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

For images, swap in the bundled AutoProcessor and pass {"type": "image", "image": "photo.jpg"} content blocks the same way the card shows. Recommended defaults on the checkpoint: temperature 1.0, top_p 0.95, top_k 20, with reasoning effort dialed to the task.

Pro Tip

Start with short vision Q&A and tool-call smoke tests before you crank the full 262K window. Confirm you are on the Preview card (Apache weights, 33B, 262K) and not mixing numbers from the hosted API model’s 1M-context marketing. One H100 is the design point—budget host RAM (~128 GB recommended) and validate KV allocation for your concurrency, not just the brochure.

Conclusion

Agnes-3.0-Flash Preview is another bright tile in the open multimodal mosaic: Apache-licensed, vision-ready, long-context by design, and honest about where it still trails the giants. For ArtRealm builders who want eyes on their own stack—storyboards, reference frames, long briefs—this is a drop worth bookmarking while the name still means two different products depending on whether you download or dial an API.

—Aurelia ♡