GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
genevera
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
genevera
Model Tree
Input Modalities
Output Modalities
Supported Functionality
| Property | Value |
|---|---|
| Architecture | Qwen3_5MoeForConditionalGeneration |
| Model Family | Qwen3.6-35B Abliterated-Heretic-Uncensored (quantized) |
| Total Parameters | ~36.2B |
| Text Hidden Size | 2,048 |
| Vision Hidden Size | 1,152 |
| Text Layers | 40 (hybrid: 3 linear + 1 full attention repeating) |
| Attention | Grouped Query Attention (GQA), 16 query heads, 2 KV heads (ratio 8:1) |
| Head Dimension | 256 |
| Vocabulary Size | 248,320 |
| Max Context Length | 262,144 tokens |
| MoE Experts | 256 per layer, 8 activated per token |
| MoE Intermediate Size | 512 |
| RoPE Theta | 10,000,000 (long-context optimized) |
| RoPE | Interleaved multi-RoPE sections [11, 11, 10] for multimodal |
| Activation | SiLU (text), GELU-Tanh (vision) |
| Transformers Version | 5.6.0.dev0 |
| Property | Value |
|---|---|
| Layers | 27 |
| Hidden Size | 1,152 |
| Patches | 16×16 |
| Temporal Patch Size | 2 |
| Spatial Merge Size | 2 |
| Output Dimension | 2,048 (matches text dimension) |
| Max Frames | 768 |
| FPS (video) | 2 |
| Property | Value |
|---|---|
| Method | AWQ (Activation-aware Weight Quantization) |
| Bits | 4-bit signed int |
| Group Size | 128 |
| Provider | auto-round v0.10.2 |
| Quantization Target | model.language_model.layers |
| Preserved (full precision) | lm_head, all shared_expert_gate layers (39), vision merger (linear_fc1, ), entire vision encoder () |
The quantization strategy preserves critical components in full precision — including the language model head, all shared expert gating mechanisms across 40 layers, the vision-language merger, and the entire vision encoder — while quantizing only the language model layer weights. This hybrid approach maintains output quality while achieving significant memory reduction.
This model supports multimodal inputs including:
<|image|> tokens (ID: 248056)<|video|> tokens (ID: 248057)<think>/</think> blocksThe model uses a sophisticated multimodal chat template with:
<thinking> / </think>)See chat_template.jinja for the complete template.
| Property | Value |
|---|---|
| Class | TokenizersBackend (HuggingFace tokenizers) |
| Processor | Qwen3VLProcessor |
| BOS / PAD | 248044 (<bos>) |
| EOS | 248046 (</s>) |
| Image Token | 248056 (`< |
| Video Token | 248057 (`< |
| Vision BOS | 248053 (`< |
See tokenizer_config.json for full configuration.
| Property | Value |
|---|---|
| Sampling | Enabled (do_sample: true) |
| Temperature | 1.0 |
| Top-K | 20 |
| Top-P | 0.95 |
| EOS Tokens | [248046, 248044] |
python
from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "./Qwen3.6-35B-A3B-Abliterated-Heretic-AWQ-4bit" # Load model and tokenizermodel = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype="auto", device_map="auto", trust_remote_code=True,)tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True) # Simple text generationmessages = [ {"role": "user", "content": "Hello, can you tell me a story?"}] text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True,)inputs = tokenizer(text, return_tensors="pt").to(model.device)outputs = model.generate(**inputs, max_new_tokens=512)print(tokenizer.decode(outputs[0], skip_special_tokens=True))python
from transformers import AutoModelForCausalLM, AutoTokenizerfrom PIL import Image model_id = "./Qwen3.6-35B-A3B-Abliterated-Heretic-AWQ-4bit" model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype="auto", device_map="auto", trust_remote_code=True,)processor = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True) # Load imageimage = Image.open("path/to/image.jpg").convert("RGB") # Multimodal conversationmessages = [ { "role": "user", "content": [ {"type": "image", "image": image}, {"type": "text", "text": "Describe this image in detail."} ] }] text = processor.apply_chat_template( messages, tokenize=False, add_generation_prompt=True,) # Note: For actual image processing, use Qwen3VLProcessor# from transformers import Qwen3VLProcessor# processor = Qwen3VLProcessor.from_pretrained(model_id) inputs = processor(text, return_tensors="pt").to(model.device)outputs = model.generate(**inputs, max_new_tokens=512)result = processor.decode(outputs[0], skip_special_tokens=True)print(result)python
from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "./Qwen3.6-35B-A3B-Abliterated-Heretic-AWQ-4bit" model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype="auto", device_map="auto", trust_remote_code=True,)tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True) # Define toolstools = [ { "type": "function", "function": { "name": "get_weather", "description": "Get the current weather for a location", "parameters": { "type": "object", "properties": { "location": {"type": "string", "description": "City name or coordinates"} }, "required": ["location"] } } }] messages = [ {"role": "system", "content": "You are a helpful assistant with weather lookup capabilities."}, {"role": "user", "content": "What's the weather in Tokyo?"}] text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, tools=tools,) inputs = tokenizer(text, return_tensors="pt").to(model.device)outputs = model.generate(**inputs, max_new_tokens=512)result = tokenizer.decode(outputs[0], skip_special_tokens=True)print(result)# The model may output a tool call in XML format:# <function=get_weather>linear_fc2model.visual.blocks| Symmetric | Yes |
| Zero Point | False |
| Version | gemm |
| Vision EOS | 248054 (`< |
| Audio BOS/EOS | Defined |
| Max Length | 262,144 tokens |