Instructions to use Blackfrost-Research/Ablagent-9B-FP8-Dynamic with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Blackfrost-Research/Ablagent-9B-FP8-Dynamic with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="Blackfrost-Research/Ablagent-9B-FP8-Dynamic")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("Blackfrost-Research/Ablagent-9B-FP8-Dynamic")
model = AutoModelForMultimodalLM.from_pretrained("Blackfrost-Research/Ablagent-9B-FP8-Dynamic", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Blackfrost-Research/Ablagent-9B-FP8-Dynamic with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Blackfrost-Research/Ablagent-9B-FP8-Dynamic"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Blackfrost-Research/Ablagent-9B-FP8-Dynamic",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/Blackfrost-Research/Ablagent-9B-FP8-Dynamic

SGLang

How to use Blackfrost-Research/Ablagent-9B-FP8-Dynamic with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Blackfrost-Research/Ablagent-9B-FP8-Dynamic" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Blackfrost-Research/Ablagent-9B-FP8-Dynamic",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Blackfrost-Research/Ablagent-9B-FP8-Dynamic" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Blackfrost-Research/Ablagent-9B-FP8-Dynamic",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use Blackfrost-Research/Ablagent-9B-FP8-Dynamic with Docker Model Runner:
```
docker model run hf.co/Blackfrost-Research/Ablagent-9B-FP8-Dynamic
```

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Ablagent-9B-FP8-Dynamic

Blackfrost Pro Series · Compact FP8 Deployment Edition

Controlled-access professional toolkit preview. This is the deployment-focused FP8 edition of Blackfrost-Research/Ablagent-9B, a compact workflow assistant for Abliteration and Domain Weight Modification work. It is not positioned as a general chat assistant, a broad coding model, or an “uncensored” release.

This repository contains a complete standalone checkpoint. It is not an adapter and it is not a requant of a third-party quantized artifact. The conversion was made directly from the exact Blackfrost BF16 release revision shown below.

Item	Value
Source	`Blackfrost-Research/Ablagent-9B`
Pinned source revision	`cf96ec74ca5fd2b2bfc78bf9af947c8d7e162173`
Architecture	`Qwen3_5ForConditionalGeneration`
Weight format	`compressed-tensors` FP8 Dynamic
Quantized weights	FP8 E4M3, channel-wise static scales
Activations	FP8 E4M3, dynamic per-token scales
Calibration corpus	None required
Configured context	262,144 tokens
FP8 payload	13,520,202,584 bytes across 3 shards
BF16 source payload	18,820,519,440 bytes across 4 shards

What Ablagent is for

Ablagent is a harness-oriented assistant for professional model-engineering workflows:

checkpoint lineage and artifact intake;
paired-corpus and prompt-provenance review;
Abliteration and DWM run preparation;
candidate comparison and evidence reconciliation;
runtime, packaging, and release diagnostics;
structured operator handoffs.

The model drafts and analyzes. The operator retains authority over tools, weight edits, deployments, publishing, and promotion decisions. Missing evidence should be reported as missing, not silently treated as passing.

FP8 conversion boundary

The release uses the Qwen3.5 FP8 Dynamic path in LLM Compressor. Selected linear layers use channel-wise FP8 weights with dynamic per-token FP8 activations. The language-model head, token embeddings, vision stack, and hybrid linear-attention modules remain in their source precision.

This is a deployment conversion only. It does not add training, DWM passes, or a new behavior claim beyond the source Ablagent checkpoint.

Shard and MTP audit

The conversion was blocked from release until the pinned BF16 source and FP8 output both passed a full index/header audit.

Check	Result
BF16 source shards present	4 / 4
Indexed source tensors accounted for	760 / 760
Output indexed tensors	888
FP8 tensors	128
Missing indexed source tensors	0

The dense 9B source declares mtp_num_hidden_layers: 1 in configuration but contains no detached mtp.* tensor payload or standalone MTP safetensors file. The official Qwen/Qwen3.5-9B Hub manifest likewise has no separate MTP file. Therefore no detached MTP payload exists to reattach; all 760 tensors present in the immediate BF16 source are represented in this release.

Machine-readable structural evidence is included in FP8_VALIDATION.json.

Runtime validation

This checkpoint was loaded with SGLang on one NVIDIA B300 as compressed-tensors at the complete 262,144-token configured context. A captured chat-completions smoke test verified the Ablagent workflow identity and basic generation (37 × 19 = 703).

Validated serving settings:

sglang serve \
  --model-path /path/to/Ablagent-9B-FP8-Dynamic \
  --served-model-name Ablagent-9B-FP8-Dynamic \
  --trust-remote-code \
  --dtype bfloat16 \
  --tp-size 1 \
  --context-length 262144 \
  --reasoning-parser mimo \
  --tool-call-parser qwen3_coder

Backend support for compressed-tensors FP8 is required. Actual concurrency and context capacity depend on GPU memory, cache precision, and serving configuration.

Lineage

Qwen/Qwen3.5-9B
XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B
Blackfrost-AI/MiMo-V2.6-Distill-Qwen-9B-Derisked-BF16
Blackfrost-Research/Ablagent-9B, epoch-3 professional workflow specialization
Ablagent-9B-FP8-Dynamic, direct deployment conversion from the pinned BF16 revision

The immediate BF16 source contains Blackfrost's earlier DWM intervention. This FP8 conversion did not perform another DWM edit.

License and use

This derivative follows the source repository's composite upstream terms. Review LICENSE and UPSTREAM-LICENSE-NOTICE.md before use. Users remain responsible for deployment controls, evaluation, applicable law, and downstream outputs.

Built by Blackfrost_AI/SI for the Blackfrost Research Pro Series.