GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Ueuee82683
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Ueuee82683
Model Tree
Input Modalities
Output Modalities
Supported Functionality
| Parameter | Value |
|---|---|
| Architecture | Pre-norm Transformer decoder (LlamaForCausalLM, arch: llama) |
| Total Parameters | 1,418,911,232 (1.42B, tied input/output embeddings) |
Layers (num_hidden_layers) | 28 |
Hidden Dimension (hidden_size) | 1536 |
FFN Dimension (intermediate_size) | 9216 (6.0x SwiGLU MLP, SiLU gate) |
| Attention Heads | 12 Query heads / 2 Key-Value heads (GQA 6:1, head_dim = 128) |
| Positional Encoding | Rotary Position Embeddings (RoPE, theta = 1,000,000.0) |
| Normalization | RMSNorm (eps = 1e-6) |
| Context Window | 32,768 tokens (8,192 default in Ollama Modelfile for low RAM usage) |
Vocabulary (vocab_size) | 49,152 custom byte-level BPE (`< |
| Weights Provided | model.safetensors (FP16, 2.71 GB) and puma-1.0-q8_0.gguf (Q8_0, 1.44 GB) |
| License | Apache-2.0 |
Puma 1.0 uses a compact 49,152-token (3 * 2^14) byte-level BPE vocabulary focused on English prose, mathematics, and source code:
0..255: Complete 256-byte UTF-8 fallback table (0 unknown tokens on any input).256..49143: 48,888 ancestor-closed BPE merges prioritized for English, symbolic math, and programming syntax.49144..49151: Dedicated control tokens (<|pad|>, <|bos|>, <|eos|>, <|system|>, <|user|>, <|assistant|>, <|fim_prefix|>, <|fim_middle|>), shrinking the embedding matrix to 75.5M parameters (49,152 x 1,536) for faster CPU token selection and lower memory bandwidth.90% Python AST validity and 70% strict unit-test pass rate on held-out algorithmic benchmarks including merge_intervals, two_sum, length_of_longest_substring, valid_brackets, binary_search, flatten, and top_k_frequent).JSON, YAML, Markdown tables) across multi-turn conversations without losing turn boundaries.~1.6 GB RAM with zero network calls.1.42B Capacity: Puma 1.0 does not store encyclopedic world trivia or obscure historical citations reliably; pair it with local document retrieval (RAG) when exact factual lookup is required.0..255), increasing token count on non-English text.bash
ollama run Merkium/puma-1.0Or run directly from Hugging Face via Ollama:
bash
ollama run hf.co/Ueuee82683/Puma-1.0Expected throughput on modern CPUs (Q8_0, 1.44 GB):
55 - 95 tokens/s20 - 40 tokens/s7 - 12 tokens/sllama.cpp CLI / Server (Pure C++, Zero Python Dependencies)Download puma-1.0-q8_0.gguf and run directly on CPU:
bash
huggingface-cli download Ueuee82683/Puma-1.0 puma-1.0-q8_0.gguf --local-dir . # Interactive multi-turn chat on CPU./llama-cli -m puma-1.0-q8_0.gguf -cnv -c 8192 -t 8 --temp 0.4 \ -p "You are Puma 1.0, an open-weight 1.42B-parameter AI assistant and coding model created by Merkium and released under Apache 2.0." # Local OpenAI-compatible HTTP API server on port 8080./llama-server -m puma-1.0-q8_0.gguf -c 8192 -t 8 --port 8080Because puma-1.0-q8_0.gguf is a standard arch: llama GGUF (1.44 GB), it runs offline on any modern smartphone with 4 GB+ RAM (iPhone 13 or newer, Pixel 7/8/9, Galaxy S22/S23/S24, or Snapdragon 8-series devices):
Ueuee82683/Puma-1.0, and download puma-1.0-q8_0.gguf (1.44 GB).<|eos|> and <|user|>, and tap Load. Expect 12 - 28 tokens/s completely offline (including in Airplane Mode).puma-1.0-q8_0.gguf directly from Ueuee82683/Puma-1.0 onto device storage and load it as a local GGUF model with context size 4096 for minimal battery drain.ollama or llama.cpp):
bash
pkg update && pkg install ollamaollama serve &ollama run Merkium/puma-1.0Once ollama run Merkium/puma-1.0 has pulled the model, query it from any Python script, IDE plugin (Continue.dev, Zed, Cursor local), or CLI tool at http://localhost:11434/v1:
python
from openai import OpenAI client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama") messages = [ {"role": "system", "content": "You are Puma 1.0, a concise local engineering assistant."}, {"role": "user", "content": "Write a Python function to merge overlapping intervals and include 2 assertions."},] response = client.chat.completions.create( model="Merkium/puma-1.0", messages=messages, temperature=0.3,)print(response.choices[0].message.content)transformers Usage (FP16 / CPU)python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "Ueuee82683/Puma-1.0"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32, device_map="auto",) messages = [ {"role": "user", "content": "Explain binary search time complexity and show a clean Python implementation."}]prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).to(model.device) stop_ids = [ tokenizer.eos_token_id, tokenizer.convert_tokens_to_ids("<|user|>"), tokenizer.convert_tokens_to_ids("<|system|>"),] with torch.no_grad(): out = model.generate( **inputs, max_new_tokens=512, temperature=0.4, top_p=0.9, do_sample=True, eos_token_id=stop_ids, pad_token_id=tokenizer.pad_token_id, ) print(tokenizer.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))Puma 1.0 uses explicit turn delimiters from its custom control vocabulary:
text
<|bos|><|system|>You are Puma 1.0, an open-weight 1.42B-parameter AI assistant and coding model created by Merkium and released under Apache 2.0.<|eos|><|user|>{user_message}<|eos|><|assistant|>{assistant_response}<|eos|>Released under the Apache License 2.0.
huggingface-cli download Ueuee82683/Puma-1.0 puma-1.0-q8_0.gguf --local-dir . # Interactive multi-turn chat on CPU./llama-cli -m puma-1.0-q8_0.gguf -cnv -c 8192 -t 8 --temp 0.4 \ -p "You are Puma 1.0, an open-weight 1.42B-parameter AI assistant and coding model created by Merkium and released under Apache 2.0." # Local OpenAI-compatible HTTP API server on port 8080./llama-server -m puma-1.0-q8_0.gguf -c 8192 -t 8 --port 8080pkg update && pkg install ollamaollama serve &ollama run Merkium/puma-1.0from openai import OpenAI client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama") messages = [ {"role": "system", "content": "You are Puma 1.0, a concise local engineering assistant."}, {"role": "user", "content": "Write a Python function to merge overlapping intervals and include 2 assertions."},] response = client.chat.completions.create( model="Merkium/puma-1.0", messages=messages, temperature=0.3,)print(response.choices[0].message.content)import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "Ueuee82683/Puma-1.0"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32, device_map="auto",) messages = [ {"role": "user", "content": "Explain binary search time complexity and show a clean Python implementation."}]prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).to(model.device) stop_ids = [ tokenizer.eos_token_id, tokenizer.convert_tokens_to_ids("<|user|>"), tokenizer.convert_tokens_to_ids("<|system|>"),] with torch.no_grad(): out = model.generate( **inputs, max_new_tokens=512, temperature=0.4, top_p=0.9, do_sample=True, eos_token_id=stop_ids, pad_token_id=tokenizer.pad_token_id, ) print(tokenizer.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))<|bos|><|system|>You are Puma 1.0, an open-weight 1.42B-parameter AI assistant and coding model created by Merkium and released under Apache 2.0.<|eos|><|user|>{user_message}<|eos|><|assistant|>{assistant_response}<|eos|>