Instructions to use trillionlabs/gravity-sft-v7-agentic with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use trillionlabs/gravity-sft-v7-agentic with Transformers:

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

pipe = pipeline("text-generation", model="trillionlabs/gravity-sft-v7-agentic")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("trillionlabs/gravity-sft-v7-agentic")
model = AutoModelForCausalLM.from_pretrained("trillionlabs/gravity-sft-v7-agentic")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use trillionlabs/gravity-sft-v7-agentic with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "trillionlabs/gravity-sft-v7-agentic"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "trillionlabs/gravity-sft-v7-agentic",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/trillionlabs/gravity-sft-v7-agentic

SGLang

How to use trillionlabs/gravity-sft-v7-agentic 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 "trillionlabs/gravity-sft-v7-agentic" \
    --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": "trillionlabs/gravity-sft-v7-agentic",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "trillionlabs/gravity-sft-v7-agentic" \
        --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": "trillionlabs/gravity-sft-v7-agentic",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use trillionlabs/gravity-sft-v7-agentic with Docker Model Runner:
```
docker model run hf.co/trillionlabs/gravity-sft-v7-agentic
```

gravity-sft-v7-agentic

Agentic SFT model trained on trillionlabs/cyan-agentic-v7. Optimized for tool-calling and terminal-style agent trajectories in three complementary formats (structured tool_calls, JSON-list, BFCL Python-syntax) with loop-recovery signal from openagent v1 augmentations.

Model summary

Architecture: DeepseekV3-style MoE (52 layers, 64 experts, top-8, 1 shared expert, 2 dense layers).
Params: ~29.5B total, ~5.3B active per token.
Context: 131,072 tokens (RoPE θ = 1e6).
Tokenizer: GLM-151k SFT.
Precision: bfloat16.

Training

Dataset: trillionlabs/cyan-agentic-v7 — 420,936 rows, mixture of xlam, nemotron (both in raw tool_calls and BFCL Python-syntax variants), envgen_v2 (TerminalBench + OfficeBench), openagent-v1 loop-recovery augmentations, and endless-terminal SFT.
Init: llama_pro_cyan_renew_lc/0000000600 (clean base checkpoint).
Epochs: 3 (step 1224 = end of epoch 3, uploaded here).
Optimizer: Muon, lr 2e-5, cosine schedule with 50-step warmup, min ratio 0.1, grad clip 1.0.
Global batch: 4.19M tokens/step (dp_shard=16, batch_size=2, seq_len=131072).
Hardware: 2× nodes with 8× B200 (16 GPUs), FSDP + activation checkpointing.
Per-turn loss masking: bad turns (looped_stuck, mistake_later_fixed, redundant_step, hallucination) are masked out of the loss via the chat template's {% generation %} markers.

Supported call formats

The model is trained to answer in three formats depending on the system prompt it receives:

Structured tool_calls — standard OpenAI-style role=assistant / tool_calls with paired role=tool observation messages.
JSON-list content — assistant emits [{"name": "fn", "arguments": {...}}] in the content field; observations come back as user turns containing a stringified list-of-dicts.
BFCL Python-syntax — assistant emits [fn_name(arg=val, ...)] and nothing else; observations come back as user turns of the form [{'role': 'tool', 'name': "fn(arg='val')", 'content': "..."}, ...].

The dataset card at trillionlabs/cyan-agentic-v7 documents the format matrix per source.

Usage

sglang

python -m sglang.launch_server \
  --model-path trillionlabs/gravity-sft-v7-agentic \
  --tp 8 \
  --dtype bfloat16 \
  --moe-runner-backend triton \
  --context-length 131072 \
  --host 0.0.0.0 --port 8080

transformers

from transformers import AutoTokenizer, AutoModelForCausalLM
tok = AutoTokenizer.from_pretrained("trillionlabs/gravity-sft-v7-agentic")
model = AutoModelForCausalLM.from_pretrained(
    "trillionlabs/gravity-sft-v7-agentic",
    torch_dtype="bfloat16",
    device_map="auto",
)

Note: the exported config.json includes moe_layer_freq: 1 so the model loads under sglang's DeepseekV3 config path without patches.

Dataset: trillionlabs/cyan-agentic-v7
Previous mix: trillionlabs/cyan-agentic-v6.1

License

Apache 2.0.

Downloads last month: -

Safetensors

Model size

30B params

Tensor type

BF16

trillionlabs
/

gravity-sft-v7-agentic