qwen3-tiny-v2/README.md

license, language, base_model, tags, model_name, pipeline_tag, library_name, quantized_by

license	language	base_model	tags	model_name	pipeline_tag	library_name	quantized_by
apache-2.0	en	Qwen/Qwen3-1.7B	qwen3 gguf layer-surgery small-language-model optimized thinking text-generation skip-connections interpolation	Qwen3-g023-tiny-v2	text-generation	llama.cpp	g023

Qwen3-g023-tiny-v2 — GGUF

An advanced 30-layer Qwen3 variant using swap, interpolation, and skip-bridge surgery.

Created through innovative layer surgery combining multi-swap, interpolation, and bridge (skip connection) techniques. Scores 94.3/100 — a 6.5-point improvement over the original Qwen3-1.7B baseline (87.8/100) and the highest score achieved in two phases of experimentation across ~250 configurations. (I have my own benchmarks, so results may vary if you run your own tests.)

Available Quantizations

Quantization	Bits/Weight	Description	Download
Q8_0	8.00	Highest quality, virtually lossless (USE THIS ONE)	Qwen3-g023-tiny-v2-Q8_0.gguf
Q6_K	6.57	Excellent quality, good compression	Qwen3-g023-tiny-v2-Q6_K.gguf
Q4_K_M	4.85	Good balance of quality and size	Qwen3-g023-tiny-v2-Q4_K_M.gguf
Q3_K_M	3.91	High compression, moderate quality loss	Qwen3-g023-tiny-v2-Q3_K_M.gguf
Q2_K	3.35	Maximum compression, significant quality loss	Qwen3-g023-tiny-v2-Q2_K.gguf

Model Details

Parameter	Value
Architecture	Qwen3ForCausalLM
Layers	30 (28 original + 2 from surgery)
Hidden Size	2,048
Intermediate Size	6,144
Attention Heads	16 query / 8 key-value (GQA)
Head Dimension	128
Vocabulary	151,936 tokens
Max Context	40,960 tokens
RoPE θ	1,000,000
Tied Embeddings	Yes
Total Parameters	~1.82B
Precision (source)	bfloat16

Surgery Operations

This model was created by applying three innovative surgical operations to Qwen/Qwen3-1.7B:

1. Multi-swap: layers 12↔13 and 16↔17 — Reorders attention blocks at two critical points in the network for improved representational flow through the mid-layers.
2. Interpolation: layers 20 & 22 (α=0.5) — Creates a new layer by blending the weights of layers 20 and 22 at equal proportions, producing a smoother transition in the upper layers.
3. Bridge (skip connection): layer 5 → after layer 20 — Copies early-layer representations (layer 5) and inserts them after layer 20, creating a skip connection that helps preserve low-level features deep in the network.

Why These Operations Work

• Multi-swap corrects suboptimal layer ordering that emerged from pre-training, allowing better gradient flow through the network's critical middle section.
• Interpolation creates a synthetic transition layer that smooths the representation gap between layers 20 and 22, reducing the information bottleneck.
• Bridge/skip connections address the "forgetting problem" in deep networks by reintroducing early feature representations at later stages — a technique inspired by ResNet's residual connections but applied at the transformer layer level.

Benchmark Results

Metric	Original (28L)	v1 (27L)	v2 (30L)	Δ vs Original
Overall Score	87.8 / 100	92.9 / 100	94.3 / 100	+6.5
Factual Accuracy	15/17 (88%)	17/17 (100%)	16/17 (94%)	+6%
Avg Perplexity	—	15.70	15.17	—
Thinking Mode	✅	✅	✅	—
Non-Thinking Mode	✅	✅	✅	—

Evaluated using a comprehensive test suite with 17 factual questions, 2 completion coherence tests, perplexity measurements, repetition analysis, and thinking/non-thinking mode verification.

Features

• Thinking mode: Full <think> / </think> reasoning support — toggle via enable_thinking parameter
• Non-thinking mode: Direct responses without chain-of-thought overhead
• Tool calling: Full function/tool calling support
• System prompts: Standard system message support
• Chat template: Qwen3 ChatML template embedded in the GGUF

Usage

With Ollama

# Download the GGUF and create from Modelfile
cat > Modelfile << 'EOF'
FROM ./Qwen3-g023-tiny-v2-Q8_0.gguf

PARAMETER temperature 1.0
PARAMETER top_p 0.95
PARAMETER top_k 45
PARAMETER min_p 0.1
PARAMETER num_ctx 40000
PARAMETER mirostat 2
PARAMETER mirostat_tau 5.0
PARAMETER mirostat_eta 0.1
PARAMETER repeat_last_n 16384
PARAMETER repeat_penalty 1.1
PARAMETER presence_penalty 0.5
PARAMETER frequency_penalty 1.0

TEMPLATE """{{- if .System }}
<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}
{{- range .Messages }}
{{- if eq .Role "user" }}
<|im_start|>user
{{ .Content }}<|im_end|>
{{- else if eq .Role "assistant" }}
<|im_start|>assistant
{{ .Content }}<|im_end|>
{{- end }}
{{- end }}
<|im_start|>assistant
"""
SYSTEM "You are a helpful assistant."
EOF

ollama create qwen3-tiny-v2 -f Modelfile
ollama run qwen3-tiny-v2

With llama.cpp

# Interactive chat
llama-cli -m Qwen3-g023-tiny-v2-Q8_0.gguf \
  --chat-template chatml -cnv

# Thinking mode
llama-cli -m Qwen3-g023-tiny-v2-Q8_0.gguf \
  -p "<|im_start|>user\nExplain quantum computing<|im_end|>\n<|im_start|>assistant\n<think>\n" \
  -n 512

# Non-thinking mode
llama-cli -m Qwen3-g023-tiny-v2-Q8_0.gguf \
  -p "<|im_start|>user\n/no_think What is 2+2?<|im_end|>\n<|im_start|>assistant\n" \
  -n 128

With Python (llama-cpp-python)

from llama_cpp import Llama

model = Llama("Qwen3-g023-tiny-v2-Q8_0.gguf", n_ctx=4096)
response = model.create_chat_completion(
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "What is the capital of France?"},
    ],
    temperature=0.6,
)
print(response["choices"][0]["message"]["content"])

System Requirements

Quantization	RAM (CPU)	VRAM (GPU)
Q8_0	~2.2 GB	~2.2 GB
Q6_K	~1.8 GB	~1.8 GB
Q4_K_M	~1.4 GB	~1.4 GB
Q3_K_M	~1.2 GB	~1.2 GB
Q2_K	~1.0 GB	~1.0 GB

v1 vs v2

This model (v2) is the Phase 2 champion, using advanced multi-operation surgery for the highest overall score.

	v1	v2 (this model)
Layers	27	30
Parameters	~1.67B	~1.82B
Operations	del + swap	swap + interpolate + bridge
Score	92.9 / 100	94.3 / 100
Factual	100% (17/17)	94% (16/17)
Perplexity	15.70	15.17
Use Case	Max factual accuracy	Max overall score

v1 is recommended when factual accuracy is paramount (100% vs 94%). v2 is recommended when overall quality matters more (94.3 vs 92.9).

Methodology

Layer surgery was performed through a systematic, test-driven process across two phases:

1. Phase 1 (~150 configs): Exhaustive search across deletion, duplication, swapping, interpolation, and combined operations → champion: del_10 + swap_11↔12 (v1)
2. Phase 2 (~95 configs): Advanced techniques including tripling, multi-swap, layer reversal, cycling, weight scaling, layer merging, skip bridges, and synthesis → champion: this model (v2)
3. Evaluation: Each configuration scored on factual accuracy (17 questions), completion coherence, perplexity, repetition ratio, and thinking mode functionality

Phase 2 Leaderboard (Top 5)

Rank	Configuration	Score	Factual	PPL
🥇	swap(12↔13,16↔17) + interp(20↔22) + bridge(5→20)	94.3	94%	15.17
🥈	swap(12↔13,16↔17) + interp(20↔22)	93.9	94%	14.74
🥉	swap(12↔13) + interp(20↔22) + bridge(5→20)	93.4	94%	15.66
4	multi-swap(12↔13,16↔17)	93.1	100%	14.90
5	Phase 1 champion (del_10 + swap_11↔12)	92.9	100%	15.70

Credits

• Base model: Qwen/Qwen3-1.7B by the Qwen team at Alibaba
• Quantization: llama.cpp
• Surgery: g023

License

Apache 2.0 — same as the original Qwen3-1.7B model.