Instructions to use AI45Research/Safin-1-35B-A3B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use AI45Research/Safin-1-35B-A3B with Transformers:

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

pipe = pipeline("text-generation", model="AI45Research/Safin-1-35B-A3B", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("AI45Research/Safin-1-35B-A3B", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use AI45Research/Safin-1-35B-A3B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "AI45Research/Safin-1-35B-A3B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AI45Research/Safin-1-35B-A3B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/AI45Research/Safin-1-35B-A3B

SGLang

How to use AI45Research/Safin-1-35B-A3B 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 "AI45Research/Safin-1-35B-A3B" \
    --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": "AI45Research/Safin-1-35B-A3B",
		"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 "AI45Research/Safin-1-35B-A3B" \
        --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": "AI45Research/Safin-1-35B-A3B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use AI45Research/Safin-1-35B-A3B with Docker Model Runner:
```
docker model run hf.co/AI45Research/Safin-1-35B-A3B
```

Safin-1 35B-A3B

Hub repository: AI45Research/Safin-1-35B-A3B. Collection: Safin-1.

This release provides inference weights and loading code. No formal current-release benchmark score table is supplied. The training-data and evaluation notes below state the known scope and remaining unknowns.

Safin-1 35B-A3B uses the MARCH architecture on a Qwen3.5 backbone (40 layers). The model type is safin1_35b_a3b. Standard persistent state is disabled.

Runtime	Prepared checkpoint value
Memory anchors	1 per 512 text tokens; learned, BOS initialized
HSD read block	32768
Global-read layers	10–30
History top-k / null gate	4 / enabled
Saved SFT step	7200
Weight inventory	17 safetensors shards, 1121 tensors

The preparation used an Intern50B-named CPT initialization followed by a NemotronSFT30B-named SFT recipe. Recorded sequence/packing length is 32768, global batch 128, learning rate 1e-5, minimum 1e-6, cosine schedule and 0.05 warmup ratio. Seed and data seed are 42. The matched baseline has the same data-list entries. Naming does not establish exact realized token counts or data redistribution rights; the companion training-data note records the known recipe.

One-step full-parameter CPT/SFT and export/fresh HF reload have been recorded on 8 GPUs, EP=8 with synthetic public data at length 1024. This does not establish convergence, optimizer-state resume, long-context quality or throughput. See the validation matrix for cached-logits and HF/vLLM limitations.

Load reviewed remote code with trust_remote_code=True. Use the actual saved runtime and chat template, and pin code/model revisions. The standard HF/vLLM launchers are under scripts/march/.

Release scope: text inference and the recorded training/runtime entrypoints. Full-file checksums are supplied. CPT corpus provenance is undocumented, and no formal benchmark score table or universal numerical-parity claim is supplied for these weights.

Do not attach internal args.json, job logs, optimizer state or private paths. The code license does not assign the weight or dataset license.

Related papers

Safin-1 is a model family trained using the MARCH architecture. This checkpoint accompanies the architecture and model-family papers:

Load from Hugging Face

Use the reviewed repository ID and a fixed commit revision. Install the recorded Python 3.11 / PyTorch 2.11 / Transformers 5.5.4 environment first; the code repository contains the complete installation recipes.

import os
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "AI45Research/Safin-1-35B-A3B"
revision = os.environ["MODEL_REVISION"]
tokenizer = AutoTokenizer.from_pretrained(model_id, revision=revision, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, revision=revision, dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
)
inputs = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Explain memory anchors."}],
    add_generation_prompt=True, enable_thinking=False,
    return_dict=True, return_tensors="pt",
).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=128, do_sample=False)
print(tokenizer.decode(outputs[0, inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

The pinned code and model configuration define the supported workload. HF Hub hosting does not itself enable the MARCH vLLM backend; accelerated vLLM inference uses the fork and launchers in the code repository. See AI45Lab/Safin-1 for software versions, training commands and numerical limitations. Use a fixed Hub commit revision for reproducibility. CPT provenance is undocumented and a formal quality score table is not supplied.

License, training data and evaluation record

The weights use Apache-2.0, as selected by the maintainer. See LICENSE and NOTICE in this model repository. Training data keeps its own terms; this does not establish permissions for the unresolved CPT corpus.

Family-level ancestry: Qwen/Qwen3.5-35B-A3B → Intern-S2-derived CPT step 5961 → SFT step 7200. The upstream training commit and CPT provenance/permission record are unavailable. The listed public dataset metadata covers the identifiable SFT source, not the entire training corpus. See TRAINING_DATA.md for the ten-file recipe and accounting limits.

See EVALUATION.md for recovered historical records and their limits. They are not a verified score table for these release payloads. The code validation matrix records separate technical acceptance and numerical limits. SHA256SUMS covers all files in this model root except the checksum file itself, including every complete weight shard.