Introduction
Introducing Nomi 2, an 4B LLM based on Qwen 3.5 4B, that thinks more efficient while keeping great performance.
In this training, we used another base and improved it's reasoning. It's base, Qwen 3.5 4B was great, but overthought most of the requests und looped.
Nomi 2 fixes that.
Even at more complicated prompts it only thinks for about 10 sec on consumer hardware (~500 Reasoning tokens).
Nomi 2 works by using one specific reasoning style:
Instead of human like CoT, Nomi uses what I like to call RASV.
In it's reasoning, it got 4 main parts:
- Restatement
- Approach
- Step-by-step derivation
- Verification
These steps help the LLM to keep the reasoning short, while keeping it effective, and avoiding loops.
🌟 Key Features & Improvements
- Architecture: Qwen-3.5-4B (runs on 8 GB VRAM GPUs like the RTX 4060).
- Multilingual Support: Can understand and generate text in German and English and many other languages.
- Efficiency: Runs with 60+ Tokens/s at Q4 with an Context Window of almost 100k tokens on consumer hardware.
📃 Evals and Showcase
Evals
Table with columns: Model Name, GSM8K*, GPQA Diamond (0-Shot)*| Model Name | GSM8K* | GPQA Diamond (0-Shot)* |
|---|
| Nomi 2.0 | 41/100 | 32/50 |
| Qwen 3.5 4B | 38/100 | 41/50 |
Reasoning Style Showcase
Q: What is 8x8x8+43+(432+432x3)-39?
Restatement
We need to evaluate the arithmetic expression 8 × 8 × 8 + 43 + (432 + 432 × 3) − 39. The operations include multiplication, addition, and subtraction, and we must apply the standard order of operations (PEMDAS/BODMAS).
Approach
Multiplication has higher precedence than addition and subtraction, so we compute all products first, then handle the parentheses grouping, then add/subtract from left to right. Specifically:
- Inside the parentheses: 432 × 3 first, then add 432
- Outside: 8 × 8 × 8, then combine everything
Step-by-step derivation
Step 1 — Inside parentheses: 432 × 3 = 1296. Then 432 + 1296 = 1728. So (432 + 432 × 3) = 1728.
Step 2 — Outside: 8 × 8 × 8 = 64 × 8 = 512.
Step 3 — Combine: 512 + 43 + 1728 − 39.
- 512 + 43 = 555
- 555 + 1728 = 2283
- 2283 − 39 = 2244
Answer: 2244.
Verification
Check by recombining: 512 + 43 + 1728 = 2283, and 2283 − 39 = 2244. Alternatively, note 512 + 1728 = 2240, then 2240 + 43 − 39 = 2240 + 4 = 2244. Both methods agree.
2244
To evaluate 8 × 8 × 8 + 43 + (432 + 432 × 3) − 39, we follow the order of operations: multiplication first, then parentheses, then addition and subtraction from left to right.
Inside the parentheses, we compute 432 × 3 = 1296, then 432 + 1296 = 1728. Outside, we compute 8 × 8 × 8 = 64 × 8 = 512. The expression now reads 512 + 43 + 1728 − 39.
Adding from left to right: 512 + 43 = 555, then 555 + 1728 = 2283, and finally 2283 − 39 = 2244.
As a sanity check, I can reorder the additions: 512 + 1728 = 2240, and 2240 + 43 − 39 = 2240 + 4 = 2244. The two independent paths give the same result, confirming the answer.
Here's my thought process for solving this math problem:
-
Analyze the Request: The user wants to solve a specific arithmetic expression: 8×8×8+43+(432+432×3)−39.
-
Identify the Operations: I need to perform multiplication, addition, and subtraction.
-
Apply Order of Operations (PEMDAS/BODMAS):
To solve the expression 8x8x8+43+(432+432x3)-39, we follow the order of operations (PEMDAS/BODMAS): Parentheses, Exponents/Multiplication, then Addition and Subtraction.
Step 1: Simplify the parentheses
Inside the parentheses, we have: (432 + 432x3)
Multiplication comes before addition.
Now the expression looks like this:
8×8×8+43+1728−39
Step 2: Perform the multiplication
Now the expression looks like this:
512+43+1728−39
Step 3: Perform addition and subtraction (left to right)
- 512+43=555
- 555+1728=2283
- 2283−39=2244
Answer:
2244
🧠 Training Details
- Base Model:
Qwen/Qwen3.5-4B
- Fine-tuning: SFT (Supervised Fine-Tuning).
- Training Tool: Unsloth (for 4-bit optimized training).
😎 Cool License
Feel free to use or improve Nomi! Benchmark results are always welcome.