GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
n4ze3m
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
n4ze3m
Model Tree
Input Modalities
Output Modalities
Supported Functionality
| Benchmark | Score | Random |
|---|---|---|
| BoolQ | 60.0 | 50.0 |
| PIQA | 58.0 | 50.0 |
| When2Call (tool-judgment MCQ) | 51.1 | 25.0 |
| WinoGrande | 50.1 | 50.0 |
| ARC-Easy | 37.9 | 25.0 |
| SocialIQA | 35.6 | 33.3 |
| HellaSwag (acc_norm) | 27.4 | 25.0 |
| OpenBookQA (acc_norm) | 26.2 | 25.0 |
| ARC-Challenge | 19.9 | 25.0 |
All via lm-evaluation-harness, 0-shot; When2Call via NVIDIA's MCQ configs adapted to this model's chat format.
Read these numbers honestly: this is a 150M model pretrained on 800M tokens (models like SmolLM2-135M use 2T). General-knowledge MCQ is its weakest suit — it is not what the model is for. The model's job is the tool-calling protocol, and there it performs:
| Behavioral probe (greedy) | Result |
|---|---|
Emits <tool_call> with valid JSON args, correct tool | ✅ |
Terminates cleanly at <|im_end|> | ✅ |
| Doesn't call tools for math/trivia/small factual questions | ✅ |
| Asks a short clarification when a required argument is missing | ✅ |
| Grounds the final answer in the tool's response | ✅ |
| Knows its name, creator, and cutoff | ✅ |
| Doesn't call tools for pure social chit-chat ("hi", "thanks") | ❌ known bug |
Pretrain 3.1h + SFT 6h + calibration ~0.6h + evals ~1h on one RTX 5090 at $0.524/h, plus a few false starts on slower boxes. Frontier labs measure training runs in GDP; this one is lunch.
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizertok = AutoTokenizer.from_pretrained("n4ze3m/assistllm-v0")model = AutoModelForCausalLM.from_pretrained("n4ze3m/assistllm-v0", dtype=torch.bfloat16, device_map="cuda")tools = [{"type": "function","function": {"name": "get_weather","description": "Get current weather for a city.","parameters": {"type": "object","properties": {"city": {"type": "string"}},"required": ["city"]},},}]messages = [{"role": "user", "content": "weather in Kochi?"}]enc = tok.apply_chat_template(messages, tools=tools, add_generation_prompt=True,tokenize=True, return_dict=True, return_tensors="pt").to("cuda")out = model.generate(**enc, max_new_tokens=80, do_sample=False,eos_token_id=tok.convert_tokens_to_ids("<|im_end|>"))print(tok.decode(out[0][enc["input_ids"].shape[1]:], skip_special_tokens=False))
text
<tool_call>{"name": "get_weather", "arguments": {"city": "Kochi"}}</tool_call><|im_end|>
Feed the tool result back as a {"role": "tool", "content": ...} message and it answers from it.
Shipped in this repo, quantized from the bf16 weights:
| File | Size | Use |
|---|---|---|
gguf/assistllm-v0-q4_k_m.gguf | 128MB | phones, browsers, Raspberry Pi |
gguf/assistllm-v0-q8_0.gguf | 172MB | near-lossless local |
gguf/assistllm-v0-bf16.gguf | 323MB | reference |
bash
llama-cli -m assistllm-v0-q4_k_m.gguf -no-cnv
⚠️ Experimental: the GGUF path (llama.cpp/Ollama) chats fine and keeps the identity, but the tool-call trigger is less reliable than the transformers path — llama.cpp's DeepSeek-MLA implementation assumes the flagship's head dims, and this model's custom dims (10 heads, nope/rope 96/32) sit slightly off that math. For dependable tool calling use the safetensors weights (above). An Ollama Modelfile with the correct ChatML Go-template is included at
ollama/Modelfile.
DeepSeek-V3-style, scaled to pocket size — 168M total / ~80M active parameters:
<tools> / <tool_call> / <tool_response>; assistant-only loss via {% generation %} blocks| Stage | Data | Tokens/rows | Time | Loss |
|---|---|---|---|---|
| A · pretrain | FineWeb-Edu + Wikipedia + Cosmopedia | 800M tokens | 3.1h | 3.31 |
| B+C · SFT | SmolTalk, general-knowledge QA, identity set, Hermes function-calling, Aquila deepsearch, bash tool, When2Call, GSM8K math, Python code | 325,659 convos | 6h | 1.87 |
| D · calibration | judgment-tilted mix: no-call / clarify / extraction + identity oversample + replay | 24,703 convos | 21min |
Stage D exists because v0.9 over-triggered: with tools in context it called them for everything, echoing the schema ({"city": "city"}) when it couldn't extract an argument — the exact pathology NVIDIA's When2Call work describes ("public datasets contain many examples of calling tools but far fewer of not calling them"). One calibration pass later: judgment fixed, identity learned, social chit-chat still trigger-happy (see below).
Model weights and code: Apache 2.0. Training data remains under its original datasets' licenses.
We leverage large language models (LLMs) to assist in drafting and polishing written content. Specifically, LLMs are used to improve clarity, coherence, and readability of the text, while ensuring that technical accuracy and intended meaning are preserved. All outputs generated by the LLM are carefully reviewed and edited by the authors to maintain factual correctness and align with the scientific content. The LLM serves as a tool to support writing efficiency, not to generate original research ideas or conclusions.
bibtex
@misc{assistllm-v0,title = {Assist-LLM v0: a 150M tool-calling model trained for less than a meal},author = {Muhammed Nazeem (n4ze3m), Page Assist},year = {2026},url = {https://huggingface.co/n4ze3m/assistllm-v0}}
| 1.20 |