GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
sifat-febo
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
sifat-febo
Model Tree
HuggingFaceTB/SmolLM2-1.7B
Input Modalities
Output Modalities
Supported Functionality
This repo ships safetensors weights and the ontor perception layer. Load it with transformers — that is the supported path.
markdown
pip install transformers torch
python
import sysfrom huggingface_hub import snapshot_downloadfrom transformers import AutoModelForCausalLM, AutoTokenizerrepo = snapshot_download("sifat-febo/banglish-companion")sys.path.insert(0, repo)from ontor.perceive import perceive, build_system_prompt, format_chatml_prompttok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo)def listen(text):prompt = format_chatml_prompt(text, build_system_prompt(perceive(text)))inputs = tok(prompt, return_tensors="pt")out = model.generate(**inputs, max_new_tokens=128, do_sample=True, temperature=0.7)return tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)print(listen("aj mood bhalo na"))
The weights file is 6.8 GB (fp32) — plan for that on download and in RAM.
The tokenizer carries no chat_template, so apply_chat_template will not work here. format_chatml_prompt builds the ChatML string the model was trained on — use it.
Your laptop. Your CPU. Nothing leaves the device.
Runs offline too. Quantized builds for llama.cpp, Ollama and LM Studio — same weights, 1.06 GB — at bangla-banglish-companion-gguf.
What's published here is safetensors. The comparison below is a record of what we did locally — not a menu of downloads.
We converted all three formats (F16, Q8_0, Q4_K_M) for our own use and ran them against BanglishBench v2.1. All three cleared it, and the spread landed inside the bench's noise — a tie. So we put the same prompts through all three side by side and compared by hand.
Q8_0 read tightest: better listening behavior, less advice-dumping than F16, without Q4's occasional inconsistency. That's the one we run ourselves.
ei model ta ki? — ekta Banglish chatbot. ekhon shune, bujhe.
ki pare? adda, khela, gaan, khabar. emotional support. safe.
ki parbe na? English/Bengali script na. choto model — sometimes bhul. doctor/lawyer na.
| Model | SmolLM2-1.7B fine-tune + ~0.5M parameter perception |
| Perception | bujhi · aabeg · bhalobasha — why → |
| Runtime | CPU. No GPU. |
| Published format | safetensors |
Evaluation results across the three formats we tested locally. All scores are measured with the perception layer attached — that is the system being evaluated, not the base weights alone.
| Format | BanglishBench v2.1 | Note |
|---|---|---|
| Q4_K_M | 97.2% | Smallest |
| Q8_0 | 96.8% | What we run locally |
| F16 | 95.7% | Closest to the published weights |
Limitations: Banglish only. May hallucinate. Not a professional advisor. Safety is best-effort (~97-100% on BanglishBench v2.1; 2-3% edge cases involve borderline deflections).
| Component | Status |
|---|---|
| Model weights — 1.7B base + perception layers | Open |
| BanglishBench v2.1 (evaluation benchmark) | Open |
Inference pipeline (ontor/perceive.py) | Open |
| Training data, pipeline, hyperparameters | Closed |
Built by Claude Code (Anthropic) under the author's direction and review. Every decision — architecture, data, what to publish, what to reject — made by the author.
Apache 2.0
bibtex
@misc{banglishcompanion2026,author = {Sifat Febo},title = {Banglish Companion: Now It Knows How You Feel},year = {2026},url = {https://huggingface.co/sifat-febo/banglish-companion}}