GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
aaronr43210
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
aaronr43210
Model Tree
Qwen/Qwen3-1.7B
Input Modalities
Output Modalities
Supported Functionality
python
from transformers import AutoModelForCausalLM, AutoTokenizerrepo = "aaronr43210/terrapulse-qwen3-1.7b"tok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto") # Prompts must match training: ChatML + an empty think block (thinking off). The runtime builds them with# backend/app/slm/prompts.py and backend/app/slm/runtime.py:render_chatml in the code repo.prompt = ("<|im_start|>system\nYou extract structured fields ... Reply with JSON only.<|im_end|>\n" "<|im_start|>user\nevening jog near Indiranagar tomorrow?<|im_end|>\n" "<|im_start|>assistant\n<think>\n\n</think>\n\n")ids = tok(prompt, return_tensors="pt").to(model.device)print(tok.decode(model.generate(**ids, max_new_tokens=200, do_sample=False)[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))The model handles two language tasks inside a deterministic system:
[E4].It holds no weather or place knowledge of its own. Dates, scores and safety limits are computed by code, and every answer is checked by a deterministic verifier (numbers + units, times, place names, citation ids and values, severity wording). An answer that fails is repaired once or replaced by a data template.
Measured bake-off on CPU (Q4_K_M, 24 hand-written questions): Qwen3-1.7B scored activity 0.917, location 1.0, profiles 0.958, scope 0.917 at 2.2 s per intent; Qwen3.5-2B scored 0.875 / 0.958 / 0.958 / 0.958 at 4.9 s. Excluded on size or licence: Phi-3/Phi-4-mini (3.8 B), Ministral-3-3B (3.85 B), Granite-4.0-micro (3.4 B), Llama-3.2-3B (3.21 B, non-OSI licence), Qwen2.5-3B (research licence), SmolLM3 (3.08 B), Gemma-3n-E2B (5.4 B).
Qwen/Qwen3-4B and kept only
because they passed the verifier, plus 16 template answers.<think></think> block, identical in training and inference.Intent extraction, 24 hand-written questions not in the training data (Q4_K_M, CPU):
| metric | base Qwen3-1.7B | TerraPulse fine-tune |
|---|---|---|
| activity | 0.917 | 1.0 |
| location | 1.0 | 0.958 |
| people profiles | 0.958 | 1.0 |
| in/out of scope | 0.917 | 0.958 |
| valid JSON | 1.0 (grammar-constrained) | 1.0 |
End to end, live data, questions never seen in training (full API with verifier):
| test | base | fine-tune |
|---|---|---|
| answer drafts passing the verifier first time (20-question set) | 4 / 13 | 12 / 14 |
| correct status (answer / clarify / decline / not available), 20 questions | 18 / 20 | 19 / 20 |
| correct status, 24 further scenarios | — | 23 / 24 (the miss was a parser bug, since fixed) |
| activity and profile extraction (44 questions) | 100 % | 100 % |
The training notebook also reported a grounding pass rate of 1.0 vs 0.55. Those 20 fact sheets were part of the training data in that run, so treat that figure as optimistic; the live first-draft numbers above are the clean ones.
Research and demonstration inside an outdoor-activity assistant that supplies live data. Health-related limits are conservative heuristics based on US EPA AQI guidance and heat-index bands; this is not medical advice.
Apache-2.0, the same as the base model. Map data © OpenStreetMap contributors; weather data by Open-Meteo.com (CC BY 4.0).
from transformers import AutoModelForCausalLM, AutoTokenizerrepo = "aaronr43210/terrapulse-qwen3-1.7b"tok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto") # Prompts must match training: ChatML + an empty think block (thinking off). The runtime builds them with# backend/app/slm/prompts.py and backend/app/slm/runtime.py:render_chatml in the code repo.prompt = ("<|im_start|>system\nYou extract structured fields ... Reply with JSON only.<|im_end|>\n" "<|im_start|>user\nevening jog near Indiranagar tomorrow?<|im_end|>\n" "<|im_start|>assistant\n<think>\n\n</think>\n\n")ids = tok(prompt, return_tensors="pt").to(model.device)print(tok.decode(model.generate(**ids, max_new_tokens=200, do_sample=False)[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))