Agnes 3.0 Flash FP8 (W8A8, calibrated)

Full FP8 quantization of Agnes-3.0-Flash Preview (33B parameters, multimodal). Weights and activations are FP8 e4m3 with static per-tensor scales, calibrated with LLM Compressor. This gives you both the model and context at FP8. Text, image, and video inputs are preserved, and the chat template is upstream's.

The Sglang patch from the original model checkpoint is included here as well.

Note that this model seems to be naturally uncensored. I ran Heretic's assessment and it gave 0/300 refusals. So no need to abliterate this model. And with that, my usual warning here:

Uncensored model: the language checkpoint has undergone abliteration to reduce refusal behavior. Treat outputs as untrusted, apply application-level safeguards, and do not assume the model will decline harmful requests.

User responsibility: this model is provided without warranty. The creators, uploaders, and maintainers are not responsible or liable for what others generate, publish, deploy, or otherwise do with this abliterated model. Users must operate it responsibly, apply appropriate safeguards, comply with applicable law, and respect third-party rights. This model is for research purposes only and is not intended for production use.

Quantization recipe

Kept at original precision: lm_head, embed_tokens, the full model.visual tower, the recurrent-state layers of delta attention (conv1d, in_proj_a, in_proj_b), and every MTP weight (shipped unquantized in model-mtp.safetensors).

agnes_quantization_manifest.json records the full build, including the SHA-256 of the calibration prompt ids used, so the run is reproducible piece for piece. recipe.yaml is the machine-readable form of the table above.

Why the quantization runs on a fused checkpoint

The upstream FFN carries two branches per layer: a main branch (width 17,408) and a parallel branch (width 2,048). This checkpoint folds them into a single set of projections per layer (intermediate_size: 19456, parallel_ffn_intermediate_size: 0).

Fusing in bf16 is concatenation and it is exact: gate/up join along the output dimension, down along the input dimension. Static per-tensor FP8 scales cannot be merged after the fact, because each branch carries its own scale and a single fused matrix has room for exactly one. Quantizing the fused layout means every serving tensor is quantized once, from the values the model will actually run, and calibration observes the same matrix geometry as inference.

Numerics

The fused layout measures 5.9e-4 full-vocabulary KL against the two-branch layout at identical weights. The residual comes from reduction order in the fused projections, and it sits below the 6.5e-4 that independent inference stacks produce from the same unquantized checkpoint. Perplexity moves 17.05 to 17.07 across the fusion.

The FP8 conversion itself has no formal before/after eval on this card. Run your own benchmarks before trusting it for a workload.

Files

The checkpoint is standard compressed-tensors float-quantized, so any engine that reads that format can load it.

Upstream and license

Derived from the Agnes-AI/Agnes-3.0-Flash Preview checkpoint at revision 891ce4f9ffb89b22888aa7fcc2bb2f3618867684, Apache 2.0, unchanged. This is a community quantization and is not affiliated with Agnes AI. The open-weight Preview differs from the production/API model, and the upstream benchmark figures below describe the Preview checkpoint before quantization.

---

The original upstream model card follows, unchanged.

Agnes-3.0-Flash Preview

Model version clarification

This repository contains an earlier open-weight Preview checkpoint of Agnes 3.0 Flash. It is distinct from the newer production/API checkpoint listed on Artificial Analysis. The Preview release has 33B parameters and a context window of 262,144 tokens. The production/API model uses a different checkpoint and configuration, with a 1M-token context window. Its benchmark results should not be attributed to the Preview weights released here. This repository was initially published as Agnes-3.0-Flash without the Preview suffix. The model card now explicitly identifies this release as Agnes-3.0-Flash Preview to clarify the distinction between the open-weight release and the production/API model. The specifications and Agnes benchmark results below refer to the Preview checkpoint. Hello! 👋 Today we are introducing Agnes-3.0-Flash Preview, an open-weights multimodal preview model built for people who want flagship-class reasoning without flagship-class hardware. Highlights:

• Competitive across core capabilities. Agnes-3.0-Flash Preview posts competitive results across reasoning, coding, and instruction-following evaluations.
• Built for demanding work. A 262 144-token context window, adjustable reasoning effort, tool calling, and text, image and video understanding.

Benchmarks

Benchmark scope: The Agnes results in the chart and table below belong to the Agnes-3.0-Flash Preview open-weight checkpoint released in this repository. They are not results for the production/API Agnes 3.0 Flash model listed on Artificial Analysis.

The Agnes-3.0-Flash Preview scores in the chart correspond to the open-weight checkpoint released in this repository. Reference results across contemporary models are shown below. The figures were compiled from different sources, harnesses, and model snapshots and do not constitute a controlled head-to-head comparison.

Higher is better for every row. Header parameter figures mix total and active counts, and harnesses and snapshot dates differ across sources, so treat cross-column comparisons as reference values rather than a controlled head-to-head evaluation.

Architecture

Agnes-3.0-Flash Preview is a hybrid-attention decoder: three of every four layers run a gated delta rule (recurrent, with per-layer state independent of sequence length), and the fourth runs standard global attention. Only 18 of the 72 layers therefore hold a KV cache that grows with context.

Quickstart

REMOTE CODE REQUIRED

Agnes-3.0-Flash Preview ships its own model implementation. Always load it with trust_remote_code=True.

Requirements

pip install "transformers>=5.12" torch torchvision accelerate

Tested on transformers 5.12.1. Image and video inputs go through the bundled processor, which needs torchvision.

Transformers

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": "请用三句话解释什么是人工智能。"}]
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))

Images and video

Image and video inputs go through the bundled processor (also remote code):

from transformers import AutoProcessor
proc = AutoProcessor.from_pretrained(path, trust_remote_code=True)
msgs = [{"role": "user", "content": [{"type": "image", "image": "photo.jpg"},
                                     {"type": "text", "text": "描述这张图。"}]}]
inputs = proc.apply_chat_template(msgs, add_generation_prompt=True, tokenize=True,
                                  return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256)
print(proc.batch_decode(out[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0])

Reasoning effort

The chat template exposes three reasoning levels — high (default), medium, low — plus a thinking-off switch:

ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt",
                              reasoning_effort="medium")   # or enable_thinking=False

Tool calling

The chat template renders tool definitions for you. The model emits calls as <tool_call><function=…><parameter=…>, and you feed results back as a tool role message:

tools = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "description": "Look up current weather for a city",
        "parameters": {
            "type": "object",
            "properties": {"city": {"type": "string", "description": "City name"}},
            "required": ["city"],
        },
    },
}]
msgs = [{"role": "user", "content": "What's the weather in Beijing right now?"}]
ids = tok.apply_chat_template(msgs, tools=tools, add_generation_prompt=True,
                              return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=256)
reply = tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True)

# <tool_call>

# <function=get_weather>

# <parameter=city>

# Beijing

# </parameter>

# </function>

# </tool_call>

# run the tool, append the result, generate the final answer
msgs += [{"role": "assistant", "content": reply},
         {"role": "tool", "content": "Clear, 26°C, light northeasterly wind"}]

Over the OpenAI API pass tools= the same way. The server returns the text above verbatim by default; to get structured tool_calls, configure sglang with a tool-call parser matching this format (likewise a reasoning parser, if you want the thinking span in reasoning_content).

SGLang

serve.sh starts a server from a stock public image, overlaying three files onto the image's sglang package and nothing else. See sglang_patch/README.md.

docker run --gpus all --shm-size 64g -p 30001:8080 \
    -v /path/to/agnes-3.0-flash:/model \
    lmsysorg/sglang:nightly-dev-20260908-20ca564b \
    bash /agnes-3.0-flash/serve.sh --served-model-name Agnes-3.0-Flash

serve.sh forwards extra command-line arguments to sglang, which is how --served-model-name takes effect; --tp 2 works the same way. The server listens on port 8080 inside the container:

from openai import OpenAI
client = OpenAI(api_key="EMPTY", base_url="http://localhost:30001/v1")
response = client.chat.completions.create(
    model="Agnes-3.0-Flash",
    messages=[{"role": "user", "content": "Design a fault-tolerant event processing architecture."}],
    temperature=1.0,
    max_tokens=2000,
)
print(response.choices[0].message.content)

Pass stream=True for streaming; tools= and reasoning_effort= are accepted the same way.

Hardware Requirements

Actual context length and concurrency depend on KV-cache allocation, runtime overhead, and tensor-parallel configuration; validate the target workload on the intended hardware.

Recommended Inference Settings

These are the checkpoint's own generation_config.json defaults.

Model Capabilities

License

Released under the Apache License 2.0.

Citation

@misc{agnes30flash2026,
  title        = {Agnes-3.0-Flash Preview},
  author       = {{Agnes AI}},
  year         = {2026},
  month        = sep,
  howpublished = {Open-weights preview checkpoint},
  url          = {https://agnes-ai.com/}
}