GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
shivrajgurjar
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
shivrajgurjar
Model Tree
Input Modalities
Output Modalities
Supported Functionality
17,408 intermediate SwiGLU neurons per layer) with zero expert-averaging or neuron dropping.64 Layers):
linear_attn) Layers: Maintain a constant-size O(1) recurrent state (72.0 MB total across all 48 layers) that never grows with sequence length.self_attn) Layers: Positioned at every 4th layer (layers 3, 7, 11, ..., 63) with 24 Query heads and only 4 Key-Value (GQA) heads (head_dim = 256).QBSC & SOSG Calibrated):
group_size = 128, packed int32 weights + float16 scales).float16 (FP16) with Quantization-Bias Spectral Compensation (QBSC) and Subnormal/Overflow Spectral Guard (SOSG).MTP) Speculative Drafter (model-00008-of-00008.safetensors):
mtp speculative decoding head for parallel multi-token verification (3.6x - 4.4x generation speedup).SEVTC Controller (arjuna_accelerators.safetensors):
SEVTC) thresholds.| File | Footprint | Description |
|---|---|---|
model-00001-of-00008.safetensors | 3.00 GiB | Layers 00-13 (Hybrid GDN + Full-Attn W4A16 + Native FP16) |
model-00002-of-00008.safetensors | 2.97 GiB | Layers 14-27 (Hybrid GDN + Full-Attn W4A16 + Native FP16) |
model-00003-of-00008.safetensors | 3.00 GiB | Layers 28-42 (Hybrid GDN + Full-Attn W4A16 + Native FP16) |
model-00004-of-00008.safetensors | 3.00 GiB | Layers 43-57 (Hybrid GDN + Full-Attn W4A16 + Native FP16) |
model-00005-of-00008.safetensors | 0.72 GiB | Layers 58-63 + Multimodal Vision Tower (visual.*) |
model-00006-of-00008.safetensors | 2.37 GiB | Token Embeddings (embed_tokens.weight) + Final Norm |
model-00007-of-00008.safetensors | 2.37 GiB | Output Unembedding Head (lm_head.weight in Native FP16) |
model-00008-of-00008.safetensors | 0.28 GiB | Native Multi-Token Prediction (mtp.*) Speculative Decoding Head |
arjuna_accelerators.safetensors | 0.16 MiB | 1M YaRN RoPE + Orthogonal Walsh-Hadamard H128,H256 & SEVTC |
| Total Model Footprint | 17.41 GiB | 2,013 Tensors Across 8 Standardized Shards |
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer repo_id = "shivrajgurjar/Arjuna-V1-1.58bit"tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained( repo_id, torch_dtype=torch.float16, device_map="auto", trust_remote_code=True,) messages = [{"role": "user", "content": "Explain how Gated DeltaNet achieves O(1) recurrent memory across 262K tokens."}]prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer(prompt, return_tensors="pt").to(model.device)output_ids = model.generate(**inputs, max_new_tokens=512)print(tokenizer.decode(output_ids[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True))import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer repo_id = "shivrajgurjar/Arjuna-V1-1.58bit"tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained( repo_id, torch_dtype=torch.float16, device_map="auto", trust_remote_code=True,) messages = [{"role": "user", "content": "Explain how Gated DeltaNet achieves O(1) recurrent memory across 262K tokens."}]prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer(prompt, return_tensors="pt").to(model.device)output_ids = model.generate(**inputs, max_new_tokens=512)print(tokenizer.decode(output_ids[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True))