Temper-2-Nodel

Temper-2-Nodel

Rust coding model of 4.2B parameters, built on Qwen3.5-4B, that answers requests in Russian and English. For Rust developers who want a small model for functions, traits, generics and threads.

Made by maylore-ai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Temper 2 Nodel is a 4.2 billion parameter language model tuned to write Rust. Among the Temper 2 models it is the biggest, and its base is Qwen3.5. It is aimed at practical Rust work with traits, generics and threads, and it understands requests in Russian and English.

Who it suits: It suits Rust developers and tool builders who want a small local model that answers coding requests in Russian or English.

Industry | Who uses it | How they use it | Example
Software | Rust developers | Ask for functions and fixes in chat form | Asking for a Fibonacci helper in Rust
Education | Rust learners | Get working examples from short, casual requests | Asking in Russian for a trait example
Tooling | Tool builders | Run a compact coding model on their own hardware | Calling it from a Python script

You can skip it if you want autocomplete inside an editor, since its training used chat requests only.

Where to use it: It runs in Python through the transformers library on Windows, macOS or Linux. A separate GGUF version exists for llama.cpp, LM Studio and Ollama.

• A Python script or notebook with transformers
• llama.cpp, LM Studio or Ollama, using the separate GGUF copy
• Chat-style use, not editor autocomplete

Use case | Who | What they get back
Write a Rust function | Developers | Working Rust code in the reply
Rust in Russian or English | Bilingual teams | Answers in either language, even to short requests
Local coding assistant | Tool builders | A small model that fits on modest hardware

How to set it up: You need Python with transformers and PyTorch. Hugging Face lists the weights at about 8.4 GB in BF16, or an estimated 2.1 GB at 4-bit.

• Python with the transformers library
• About 8.4 GB of disk for the BF16 weights
• A GPU is handy, since the code lets the library choose the device

1. Install the Hugging Face download tool.
2. Download the model files, or let transformers fetch them by name.
3. Load the model and tokenizer with the script under How to use it.

pip install -U huggingface_hub
hf download maylore-ai/Temper-2-Nodel

How to use it: Send a chat message through the tokenizer's chat template and generate up to 1000 new tokens. The template adds a Temper system prompt when you give none, and the model answers straight away with no visible reasoning. The recommended temperature is 0.2.

from transformers import AutoModelForCausalLM, AutoTokenizer

name = "maylore-ai/Temper-2-Nodel"
tokenizer = AutoTokenizer.from_pretrained(name)
model = AutoModelForCausalLM.from_pretrained(name, dtype="auto", device_map="auto")

messages = [{"role": "user", "content": "Write a Rust function that returns the n-th Fibonacci number."}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=1000)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Tool | What it does
AutoTokenizer | Loads the tokenizer and its chat template
AutoModelForCausalLM | Loads the model and places it on your device
generate | Writes the reply, here up to 1000 new tokens

Example: You ask it for a Rust function that gives the n-th Fibonacci number. The model replies right away with Rust code for it, and the script prints only the newly written text.

For newcomers and experts: If you are new, run the script above and read the Rust it prints. Try changing the request to a different small function and compare the answers.

If you know the field, it is Qwen3.5-4B fine-tuned on Rust tasks and dialogs plus hard problems, then trained with reinforcement learning on tasks with tests, then merged with the original weights. Its answers became less varied, so it trails the base on MBPP-RS at ten attempts.

Qwen3.5-4B base
Fine-tune on Rust tasks and dialogs
Reinforcement learning on tasks with tests
Merge with the original weights

Good to know: On first-try HumanEval-RS, Qwen3.5-9B is ahead by 2.8 points. The model often explains its own code poorly.

• License: Apache 2.0
• Cost to run: no API key, account or paid service is named, but the model needs memory for about 8.4 GB of weights
• Data it reaches: your prompts on your own machine, plus files downloaded from Hugging Face
• Not for: inline editor completion, or reliable explanations of the code it writes

Title: Temper-2-Nodel scores, in percent
HumanEval-RS pass@1: 73.2%
MBPP-RS pass@1: 64.8%
Temper Bench v2: 37.3%

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 4.2B
• Weights: about 8.4 GB in BF16; about 2.1 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 451
• Likes: 3
• License: apache-2.0
• Languages: en, ru
• Based on: Qwen/Qwen3.5-4B
• Quantized versions: 2
• Tags: transformers, safetensors, qwen3_5_text, text-generation, rust, code, qwen3_5, conversational

Links

• Model: https://huggingface.co/maylore-ai/Temper-2-Nodel
• Files: https://huggingface.co/maylore-ai/Temper-2-Nodel/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download maylore-ai/Temper-2-Nodel

Then follow the usage in its model card. Nothing in this kit runs by itself.