Hy-Embodied-VLM-1.0-Text-Only

Hy-Embodied-VLM-1.0 Text-Only

This is the text-generation backbone extracted from tencent/Hy-Embodied-VLM-1.0 at source revision d26cbcc2a7f3e2e4e9fd93398645638465ba5581.

It contains the 48-layer Hy3-A3B Mixture-of-Experts causal language model and does not contain the Hy-ViT2 vision encoder. It is a deterministic weight extraction, not a fine-tune. Image and video inputs are not supported.

Model details

Usage

The official custom model code is bundled, so trust_remote_code=True is required. The versions used for structural validation were transformers==4.57.6 and torch==2.9.1.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "sasa2000/Hy-Embodied-VLM-1.0-Text-Only"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    device_map="auto",
)

messages = [{"role": "user", "content": "Explain how to safely open a refrigerator."}]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    enable_thinking=True,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=256)

print(tokenizer.decode(output[0], skip_special_tokens=False))

Set enable_thinking=False for direct-answer mode. Only text content should be passed to the chat template; the image placeholders from the source VLM are not usable in this checkpoint.

The raw text weights occupy about 60.13 GB (56.00 GiB). Loading requires additional RAM or VRAM for framework overhead, caches, and generation state.

Extraction

The source checkpoint contains 19,117 tensors. This conversion:

• keeps 18,673 language-model tensors;
• removes all 444 vit.* tensors;
• removes multimodal processor configuration;
• retains the original tokenizer and enable_thinking chat template;
• preserves every retained tensor's dtype, shape, and raw data bytes.

The released checkpoint names and released custom model class use different names for 143 otherwise equivalent parameters. The following key-only normalizations make the checkpoint loadable by the bundled HYV3VLForCausalLM implementation:

Equivalent old/new MoE configuration aliases required by the released remote code are also included: moe_intermediate_size, moe_topk, num_shared_expert, moe_layer_num_skipped, and use_qk_norm.

Validation

The following checks passed locally:

• AutoConfig loads as HYV3VLConfig with architecture HYV3VLForCausalLM;
• tokenizer size equals the configured vocabulary size (120818);
• the text-only chat template renders in thinking and non-thinking modes;
• all safetensors headers, offsets, file sizes, dtypes, shapes, and index entries are structurally consistent;
• no vision, image, video, or projector tensor names remain;
• a meta-device HYV3VLForCausalLM exposes exactly 18,673 state-dict keys, matching the checkpoint index with zero missing and zero extra keys.

See validation_report.json and extraction_manifest.json for machine-readable details.

Limitations

• This checkpoint supports text generation only. It cannot accept images or videos and does not reproduce the source model's multimodal behavior.
• It uses custom Transformers code and therefore requires trust_remote_code=True. Review the bundled Python files before loading in a sensitive environment.
• Capabilities, intended use, safety considerations, and inherited limitations should be understood in the context of the original model card.

License and attribution

The source model is released under the Apache License 2.0. This derivative checkpoint retains the source LICENSE. Please cite and attribute the original Tencent Hy-Embodied-VLM-1.0 release when using this model.