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

Libraries

How to use Blackfrost-Research/Ablagent-9B 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")
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")
model = AutoModelForMultimodalLM.from_pretrained("Blackfrost-Research/Ablagent-9B", 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 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"
# 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",
		"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

SGLang

How to use Blackfrost-Research/Ablagent-9B 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" \
    --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",
		"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" \
        --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",
		"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 with Docker Model Runner:
```
docker model run hf.co/Blackfrost-Research/Ablagent-9B
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

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.

Blackfrost Pro Series

Ablagent-9B · Compact Model Engineering Toolkit

Controlled-access toolkit preview. Ablagent-9B is a compact, task-specific workflow model designed to operate with the bounded Blackfrost Ablagent Standalone Harness. It is built for professional model-engineering work—not as a general chat assistant, broad coding model, or autonomous weight-editing system.

The Pro Series

Blackfrost Pro Series is a line of specialized models packaged as professional tools and workflows. Each release is designed around a defined operational role instead of open-ended chat.

The series format combines:

a domain-tuned model;
a deliberate operating boundary;
a harness or integration surface;
deployment guidance;
a validation and provenance record.

This structure is intended to extend across model engineering, marketing operations, logistics, and other bounded professional workflows while preserving a recognizable Blackfrost product and documentation standard.

Pro Series standard	What it means
Task-specific	Every model has an explicit professional role and workflow
Toolkit-first	The model is paired with an operating surface, not presented as a bare chatbot
Private-deployment ready	Local-first operation and controlled endpoints are first-class concerns
Operator-governed	Tools, data access, write authority, and promotion decisions remain explicit
Evidence-led	Measured results are separated from assumptions, impressions, and untested claims

Product brief

Ablagent-9B is the compact Blackfrost Pro Series toolkit for Abliteration and Domain Weight Modification workflows. It helps an operator inspect artifacts, maintain experimental discipline, prepare structured plans, compare candidate evidence, diagnose incomplete runs, and produce clear handoff records.

The model is intended to be paired with Blackfrost Ablagent Standalone Harness 1.0.0, which turns the checkpoint into a bounded, local-first workflow assistant. The current harness is a read-only evaluation preview: it can inspect an authorized workspace through a small tool surface, but it cannot execute shell commands, modify weights, deploy services, publish artifacts, or write into the workspace.

Toolkit profile


Series	Blackfrost Pro Series
Discipline	Model Engineering
Toolkit	Ablagent
Model	Ablagent-9B
Primary workflow	Abliteration and Domain Weight Modification
Operating mode	Evidence-focused, operator-directed, harness-bounded
Current tool boundary	Read-only workspace inspection
Delivery	Full BF16 checkpoint + companion standalone harness
Release status	Controlled-access professional toolkit preview

Professional workflow

Ablagent is tuned to assist with the work surrounding a controlled weight-modification program:

Provenance and intake — checkpoint archaeology, lineage, revision, and source-fingerprint review;
Capture preparation — paired-corpus hygiene, role leakage, template leakage, and prompt provenance checks;
Direction and edit review — capture interpretation, edit-surface enumeration, target-set review, and MTP preservation discipline;
Candidate evaluation — geometry, distribution drift, behavioral evidence, and HITL comparisons across alpha/pass candidates;
Run recovery — reconcile manifests, receipts, controls, and filesystem evidence without silently inventing missing state;
Release operations — runtime diagnosis, structured tool-use checks, packaging review, and evidence-focused handoff records.

The intended output is not merely an answer. It is a more legible, auditable workflow for the human operator making the decision.

Operating boundary

Ablagent does not receive access to files, weights, GPUs, endpoints, or external systems unless an operator explicitly connects them.
The model may draft plans, diagnostics, scripts, and reports, but those outputs require review.
Structural correctness, behavior, distribution drift, and operator acceptance remain separate gates.
Missing measurements are recorded as missing; they are not treated as zero or as passing.
A served candidate is not automatically an accepted or promoted checkpoint.

What this product is not

not a general-purpose chat or coding assistant;
not positioned as a replacement for larger open-domain models;
not an autonomous DWM engine;
not a substitute for the actual checkpoint, tensor inventory, manifests, or measurements;
not a claim that every backend, context length, tool schema, or task has been validated;
not a blanket claim about refusal behavior.

Why 9B

The smaller checkpoint is intentional. Ablagent-9B is practical to host privately, responsive during iterative lab work, and economical enough to keep beside an active experiment. It was selected as a focused model-engineering sidecar—not to compete with much larger general-purpose models on broad knowledge, coding, or benchmark coverage.

Ablagent edition	Recommended role
Ablagent-9B	Compact sidecar for routine evidence triage, run-state inspection, structured reporting, and bounded harness tasks
Ablagent-35B-A3B	Pro Series flagship for denser comparisons and more involved model-engineering workflows

Both editions are workflow assistants operating inside the same professional boundary.

Technical profile


Architecture	`Qwen3_5ForConditionalGeneration` / Qwen3.5 hybrid linear + full attention
Parameters	9,409,813,744
Precision	BF16
Packaging	Full standalone merged checkpoint; no adapter required
Weight payload	18,820,427,744 bytes across four safetensors shards
Text stack	32 layers, hidden size 4,096, intermediate size 12,288
Configured context	262,144 tokens
Training selection	Epoch 3, global step 375
Immediate base	`Blackfrost-AI/MiMo-V2.6-Distill-Qwen-9B-Derisked-BF16`
Base revision	`a99b2f3cd25382732bfce6e6022eca2dae982a96`
Reasoning parser	`mimo`
Tool-call parser	`qwen3_coder`
Validated runtime	SGLang on one NVIDIA B300
Validated image	`lmsysorg/sglang@sha256:06e4f2ed21afde4ff513cda65070124e727ba23ccaeff7712b8c40e1097d611f`

The Qwen3.5 conditional-generation architecture includes inherited vision components. This specialization and the published evaluation evidence are text-focused; multimodal capability was not re-benchmarked for this release.

Lineage

Qwen/Qwen3.5-9B
XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B, pinned upstream revision 2367e865d009c13ac81713a2878291d33ab28177
Blackfrost-AI/MiMo-V2.6-Distill-Qwen-9B-Derisked-BF16, a Blackfrost DWM derivative at revision a99b2f3cd25382732bfce6e6022eca2dae982a96
Ablagent-9B, an epoch-3 supervised specialization merged into a full BF16 checkpoint

The immediate base already contains Blackfrost's DWM intervention. No new DWM weight edit was performed during the Ablagent fine-tuning stage. Ablagent's DWM identity describes its specialist purpose; it must not be read as a claim that this epoch introduced or independently revalidated a new DWM direction.

The specialist identity is carried by the tuned checkpoint and its packaged deployment template:

You are Ablagent, an AI assistant for Abliteration and DWM workflows, built by Blackfrost_AI/SI.

Training record

The selected checkpoint completed three supervised passes over a curated internal reasoning set:

Item	Recorded value
Training examples	500
Validation examples	40
Covered domains	20
Final global step	375
Input tokens seen	2,751,052
Final train loss	1.3108698120
Epoch-3 validation loss	1.3625452518

The data build required exact system → user → assistant records, removed source system prompts, inserted the Ablagent identity, rejected speaker/template leakage, and used assistant-only labels. The label audit found no user marker in the supervised sample.

The private corpus, raw examples, tokenized data, LoRA adapter, optimizer state, training scripts, and training recipe are not included in this repository.

Why epoch 3

Checkpoint	Validation loss
Epoch 1	1.3852348328
Epoch 2	1.3647842407
Epoch 3	1.3625452518

Epoch 3 was selected by the operator after side-by-side review. The validation-loss improvement over epoch 2 is small and should not be interpreted as a broad capability result.

Evaluation record

Identity smoke

The validated epoch-3 serve identified itself as Ablagent, attributed Blackfrost_AI/SI, and stated that it assists with Abliteration and DWM workflows.

Five-prompt reasoning snapshot

The same five prompts were run against the immediate BF16 base and epoch 3 with temperature 0, thinking enabled, and a 4,096-token completion cap.

Measure	Base BF16	Ablagent epoch 3
Prompts	5	5
Responses with captured reasoning	4	5
Normal stops	4	4
Total completion tokens	7,563	6,920
Total reasoning tokens	6,439	5,721

Manual review found correct algebra and exact-JSON answers plus a plausible causal diagnosis. It also found two material failures: one logic-scheduling response exhausted its reasoning budget without a final answer, and one Python-debugging response proposed an invalid fallback for unhashable values. The Python result remained incorrect when rerun with an 8,192-token cap.

This is a small qualitative snapshot, not a benchmark claim. General capability, tool-use, multimodal quality, long-context quality, and the fine-tuned derivative's full-vocabulary distribution drift remain incompletely measured.

DWM workflow posture

Ablagent is trained to keep several distinctions explicit:

structural correctness is not behavioral quality;
separation reduction is not sufficient evidence for promotion;
source, capture prompt, serving prompt, tokenizer, and template are separate variables;
automated gates and operator HITL are separate decisions;
missing measurements are recorded as missing, never treated as zero;
MTP/draft tensors are donor state and must remain outside DWM edit surfaces;
a served candidate is not automatically an accepted checkpoint.

These are intended reasoning habits, not guarantees. Operators should use the actual DWM measurement bundle and validator as the source of truth.

Included toolkit

Model checkpoint

complete merged epoch-3 BF16 weights;
Ablagent identity carried by the tuned checkpoint and packaged deployment template;
separate reasoning and final-answer fields on the validated serving stack;
OpenAI-compatible Chat Completions deployment target;
no adapter required at inference time.

Companion standalone harness

Ablagent-9B is designed to be paired with Blackfrost Ablagent Standalone Harness 1.0.0, distributed with the Ablagent-35B-A3B release. The same harness supports both Ablagent editions.

It is a local-first, read-only evaluation preview with an OpenAI-compatible client, workspace boundary, staged-plan driver, receipt checks, transcript controls, and the read_file, grep, and glob tools. It exposes no shell, write, deployment, publication, or weight-editing tools.

That restriction is part of the product definition, not a missing feature in the model card. The harness presents Ablagent as a focused workflow assistant rather than an open-ended chatbot. Set the harness model field to the exact identifier returned by the serving endpoint. The canonical served model ID for this checkpoint is Ablagent-9B.

Deployment

The complete sanitized deployment kit is in DEPLOYMENT/.

Download

hf download Blackfrost-AI/Ablagent-9B --local-dir ./Ablagent-9B
cd Ablagent-9B
sha256sum -c SHA256SUMS

Native SGLang

MODEL_PATH="$PWD" bash DEPLOYMENT/serve-native.sh
python3 DEPLOYMENT/smoke-test.py --base-url http://127.0.0.1:8000/v1

Docker

MODEL_DIR="$PWD" bash DEPLOYMENT/serve-docker.sh
python3 DEPLOYMENT/smoke-test.py --base-url http://127.0.0.1:8000/v1

The scripts default to loopback binding and one GPU. Review the documented environment variables before changing host exposure, context, GPU selection, or concurrency.

Serving contract

The validated server pairing uses:

OpenAI-compatible Chat Completions;
served model name Ablagent-9B;
mimo reasoning parser;
qwen3_coder tool-call parser;
reasoning_content kept separate from final content;
tensor parallel size 1 on the validated B300 deployment.

The model files preserve the inherited MiMo template. The deployment kit additionally uses DEPLOYMENT/chat_template_ablagent.jinja, which injects only the Ablagent identity; it does not embed a DWM recipe or operational policy. Preserve the packaged template and allow enough completion tokens for both reasoning and the final answer. Tool schemas, reasoning parser, tool-call parser, backend, and template should be validated as one serving contract.

Limitations and operator responsibilities

Ablagent-9B is a compact specialist and is not expected to match larger models on broad, open-domain capability.
This release has not completed a broad benchmark suite.
The 262,144-token configuration is not a claim of full-length quality or universal hardware fit.
The DWM-specific full-distribution measurement suite was not rerun after the epoch-3 fine-tune.
Tool calling was configured with the known-good parsers, but this release does not claim a comprehensive multi-step tool battery.
Outputs can be incorrect, incomplete, overconfident, or based on misunderstood tensor/runtime evidence. Review before executing weight edits or destructive operations.
The current harness is a read-only evaluation preview, not a production automation controller.
Operators remain responsible for authorization, access control, data handling, validation, and destructive-action review, and lawful deployment.

Repository contents

Full epoch-3 BF16 model weights
Configuration, tokenizer, processors, and inherited model chat template
Blackfrost Pro Series model card and banner
Sanitized SGLang native/Docker deployment kit and smoke test
Compatibility with Blackfrost Ablagent Standalone Harness 1.0.0, distributed with the companion Ablagent-35B-A3B release
Upstream license notice and SHA-256 manifest

This repository intentionally excludes training data, raw traces, private prompts, training and DWM recipes, direction banks, direction vectors, adapters, optimizer/checkpoint state, internal measurement bundles, credentials, logs, and live infrastructure addresses.

License and attribution

The immediate Xiaomi MiMo derivative declares MIT metadata, while the underlying Qwen3.5-9B repository ships Apache-2.0 terms. See UPSTREAM-LICENSE-NOTICE.md and the bundled LICENSE for the carried upstream notice. Downstream users are responsible for reviewing all applicable upstream terms.

BLACKFROST PRO SERIES
Professional models. Defined workflows. Operator control.

@Blackfrost_AI

Blackfrost_AI/SI · Model engineering and professional workflow systems