GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
jmarceno
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
jmarceno
Model Tree
google/gemma-4-12B-it-qat-q4_0-unquantized
Input Modalities
Output Modalities
Supported Functionality
Scored on a 128-task executable harness. Each task asks for one Elixir module. A pass means the completion parses, compiles, and its tests pass on Elixir 1.20.2 / OTP 29. One completion per task, temperature 0, 16,384-token context, llama.cpp. The base and the fine-tune were both scored as 4-bit GGUF under that same setup.
| Gemma 4 12B QAT | Elixir Gemma 4 12B | Change | |
|---|---|---|---|
| Tasks passed | 31 / 128 | 46 / 128 | +15 |
| Pass rate | 24.2% | 35.9% | +11.7 pp |
| Relative gain | +48% |
The fine-tune solves about half again as many Elixir tasks as the base model it was trained from.
| Area | Base | Elixir Gemma 4 12B |
|---|---|---|
| Standard library | 6 / 16 | 10 / 16 |
| Error handling | 8 / 16 | 11 / 16 |
| OTP | 2 / 16 | 4 / 16 |
| BEAM | 1 / 16 | 3 / 16 |
| Abstractions | 1 / 16 | 3 / 16 |
Completions are also cleaner: warning-free rate moves from 46.9% to 58.6%, and every completion stays inside a single fenced answer.
bash
llama-server -m Elixir-Gemma-4-12B-Q4_K_M.gguf -c 16384 --temp 1.0 --top-k 64 --top-p 0.95The benchmark file is Elixir-Gemma-4-12B-Q4_K_M.gguf.
Needs a Transformers build that knows gemma4_unified.
python
import torchfrom transformers import AutoTokenizer, Gemma4UnifiedForConditionalGeneration model_id = "jmarceno/Elixir-Gemma-4-12B"tokenizer = AutoTokenizer.from_pretrained(model_id)model = Gemma4UnifiedForConditionalGeneration.from_pretrained( model_id, dtype=torch.bfloat16, device_map="auto",) messages = [ { "role": "user", "content": "Write an Elixir module Counter with a GenServer that stores an integer and supports increment and get.", }]prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer(prompt, return_tensors="pt").to(model.device)output = model.generate(**inputs, max_new_tokens=512)print(tokenizer.decode(output[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))Apache 2.0, the same license as google/gemma-4-12B-it-qat-q4_0-unquantized.
llama-server -m Elixir-Gemma-4-12B-Q4_K_M.gguf -c 16384 --temp 1.0 --top-k 64 --top-p 0.95import torchfrom transformers import AutoTokenizer, Gemma4UnifiedForConditionalGeneration model_id = "jmarceno/Elixir-Gemma-4-12B"tokenizer = AutoTokenizer.from_pretrained(model_id)model = Gemma4UnifiedForConditionalGeneration.from_pretrained( model_id, dtype=torch.bfloat16, device_map="auto",) messages = [ { "role": "user", "content": "Write an Elixir module Counter with a GenServer that stores an integer and supports increment and get.", }]prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer(prompt, return_tensors="pt").to(model.device)output = model.generate(**inputs, max_new_tokens=512)print(tokenizer.decode(output[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))