GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Erog0291
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Erog0291
Model Tree
Input Modalities
Output Modalities
Supported Functionality
We ran StewMind 1.7B live on a plain CPU (no GPU) across 5 agentic tasks. Raw results are published in benchmark_results.json.
Tool-routing test — input: "Send 5000 naira to my brother today. Reply with only a JSON object." StewMind answered:
json
{"tool": "transfer", "args": {"amount": "5000", "recipient": "my brother", "date": "today"}}Valid JSON. Correct tool. Correct arguments. On a CPU.
Measured on CPU (fp32, unquantized, shared sandbox):
| Test | Result |
|---|---|
| Intent routing → JSON tool call | ✅ valid JSON, correct tool + args |
| Tool call (weather) | ✅ correct plan |
| ReAct planning | ✅ sound first step |
| Memory extraction | ✅ correct approach |
| Pidgin understanding | ✅ understood "Wetin be your main work?" |
| Speed | ~2 tok/s on plain CPU fp32 (15-30+ tok/s typical with GGUF Q4 on a laptop) |
| Footprint | 1.72B params · 1.0 GB GGUF (Q4_K_M) · loads in seconds |
| StewMind 1.7B | GPT-4o-mini (API) | Llama-3.1-8B | Phi-3.5-mini | |
|---|---|---|---|---|
| Runs offline | ✅ 100% local | ❌ needs internet | ✅ | ✅ |
| Data leaves your device | Never | Always | Never | Never |
| Download size | ~1 GB (GGUF) | 0 (API only) | ~4.7 GB (Q4) | ~2.3 GB (Q4) |
| RAM to run | ~1.5 GB | n/a | ~6 GB | ~3 GB |
| Cost per 1M tokens | $0.00 forever | ~$0.15–0.60 (listed) | $0.00 | $0.00 |
| First-token latency | ms (on-device) | ~500-2000 ms network | ms | ms |
| Fits in a phone | ✅ | ❌ | ❌ | barely |
A 1.7B will not out-write GPT-4. It doesn't have to. Agents need routing, tool calls and reflexes — and there StewMind wins on speed, cost and privacy. NVIDIA's own research ("Small Language Models are the Future of Agentic AI") points exactly this way.
python
from transformers import AutoModelForCausalLM, AutoTokenizertok = AutoTokenizer.from_pretrained("Erog0291/stewmind-1.7b")model = AutoModelForCausalLM.from_pretrained("Erog0291/stewmind-1.7b")msgs = [{"role": "user", "content": "Call the weather tool for Lagos. JSON only."}]ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt")print(tok.decode(model.generate(ids, max_new_tokens=80, do_sample=False)[0][ids.shape[1]:], skip_special_tokens=True))GGUF (Ollama / llama.cpp): download stewmind-1.7b-Q4_K_M.gguf from this repo — ready to run on phones, laptops and edge devices.
ULTRA is training right now on 100,000 agentic workflow samples (tool calls, ReAct traces, memory ops) — checkpoints land in Erog0291/stewmind-1.7b-ultra as they bake. Expect sharper, faster, stricter-JSON agents.
Read the launch story: Big Models Cook for Everyone. This One Cooks for You.
Built with 🍲 in Nigeria by MUTYINT · Apache 2.0 · commercial use allowed
{"tool": "transfer", "args": {"amount": "5000", "recipient": "my brother", "date": "today"}}from transformers import AutoModelForCausalLM, AutoTokenizertok = AutoTokenizer.from_pretrained("Erog0291/stewmind-1.7b")model = AutoModelForCausalLM.from_pretrained("Erog0291/stewmind-1.7b")msgs = [{"role": "user", "content": "Call the weather tool for Lagos. JSON only."}]ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt")print(tok.decode(model.generate(ids, max_new_tokens=80, do_sample=False)[0][ids.shape[1]:], skip_special_tokens=True))