GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
vivekchakraverty
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
vivekchakraverty
Model Tree
Qwen/Qwen3-1.7B
Input Modalities
Output Modalities
Supported Functionality
Given a structured brand intake (name, audience, archetype, competitors, personality sliders, channels, …) it writes one deliverable at a time from a fixed set of 12:
| Phase | Sections |
|---|---|
| Strategy | Brand Core, Positioning Statement, Audience Personas, Competitive Frame, Archetype Expression |
| Verbal Identity | Brand Personality & Voice, Messaging Pillars, Tagline Territory |
| Visual Identity | Visual Direction Brief |
| Experience | Channel Voice Adaptation, Brand Guardrails |
| Growth | 90-Day Activation Checklist |
It was trained on one exact prompt shape, and it degrades noticeably if you depart from it. Use the chat template with thinking disabled.
System message:
markdown
You are a senior brand strategist writing one deliverable for a client's Brand Document.Be concrete, specific, and free of marketing filler. Output clean markdown for a singledeliverable (do not repeat the whole document). Never invent brand facts beyond what theintake provides — if information needed for a claim is missing, write 'TBD — needs founderinput' instead of inventing it.
User message:
markdown
DELIVERABLE: {section name}INSTRUCTIONS:{section instructions}BRAND INTAKE (ground truth — do not contradict, do not invent beyond this):{intake as JSON, indent=2}Write only the "{section name}" deliverable now, in markdown.
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizermodel_id = "vivekchakraverty/qwen3-1.7b-brandforge"tok = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.float16).to("cuda")prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)inputs = tok(prompt, return_tensors="pt").to("cuda")out = model.generate(**inputs, max_new_tokens=1500, do_sample=True, temperature=0.7, top_p=0.9)
Qwen3 emits an empty <think></think> block even with thinking disabled — strip
it before showing the text to a user.
Supervised fine-tuning followed by DPO, on brand-document sections produced by a larger teacher model and then put through a critique-and-revise pass and a judge filter. Retrieval was deliberately discarded at training time: the branding knowledge is meant to live in the weights so that inference needs nothing but the prompt above.
TBD — needs founder input rather than invent facts,
but it can still overreach — check any claim that reads like a statistic,
certification, or customer result.Apache-2.0, inherited from Qwen3-1.7B.