Skip to content

Qwen3.5 Model Comparison: 27B vs 35B-A3B

Date: 2026-02-25 Hardware: RTX 4090 (24GB VRAM) Test: Multi-agent Tetris development (Planner → Developer → QA)

Models Under Test

Model Preset Quant Port VRAM Parallel
Qwen3.5-27B qwen35-27b-multi Q4_K_XL 7082 17 GB 3 slots
Qwen3.5-35B-A3B qwen35-35b-q3-multi Q3_K_XL 7081 16 GB 3 slots
Qwen3.5-35B-A3B qwen35-35b-multi Q4_K_XL 7080 20 GB 3 slots

Architecture comparison: - 27B: Dense model, 27B total / 27B active params - 35B-A3B: Sparse MoE, 35B total / 3B active params

Charts

Total Time Comparison

Total Time

Phase Breakdown

Phase Breakdown

VRAM Efficiency

VRAM Efficiency

Code Output Comparison

Code Output

Results

Summary

Model VRAM Total Time Plan Dev QA Lines Valid
Qwen3.5-27B Q4 17 GB 134.0s 36.3s 72.1s 25.6s 312 YES
Qwen3.5-35B-A3B Q3 16 GB 34.8s 7.3s 20.1s 7.5s 322 YES
Qwen3.5-35B-A3B Q4 20 GB 37.8s 8.2s 22.0s 7.6s 311 YES

Key Findings

  1. 35B-A3B models are dramatically faster than 27B — 35s vs 134s (3.8x faster!)
  2. 35B-A3B Q3 is fastest overall — 34.8s total, uses only 16GB VRAM
  3. 35B-A3B Q4 slightly slower than Q3 — 37.8s vs 34.8s (8% slower, 4GB more VRAM)
  4. 27B is surprisingly slow — Dense architecture less efficient than sparse MoE
  5. All models produced valid, runnable code — 311-322 lines each

Speed Comparison

Phase 27B Q4 35B-A3B Q3 35B-A3B Q4 35B-A3B Q3 vs 27B
Planning 36.3s 7.3s 8.2s 5.0x faster
Development 72.1s 20.1s 22.0s 3.6x faster
QA Review 25.6s 7.5s 7.6s 3.4x faster
Total 134.0s 34.8s 37.8s 3.8x faster

VRAM Efficiency

Model VRAM Time VRAM Efficiency
35B-A3B Q3 16 GB 34.8s Best (fastest, lowest VRAM)
27B Q4 17 GB 134.0s Worst (slow, mid VRAM)
35B-A3B Q4 20 GB 37.8s Good (fast, highest VRAM)

Generated Code & QA Analysis

All three models produced functional Tetris games with similar structure:

Model Lines Chars Syntax QA Verdict
27B Q4 312 11,279 VALID Issues noted
35B-A3B Q3 322 11,260 VALID Issues noted
35B-A3B Q4 311 10,260 VALID Issues noted

QA Review Summary

All three QA agents identified similar potential issues in the generated code:

Common observations across models: - Collision detection edge cases (pieces near board edges) - Rotation wall-kick not fully implemented - Score calculation could have edge cases with >4 lines - Game over detection timing

Verdict: All three games compile and run correctly. The QA agents were thorough in identifying potential edge cases, but the core gameplay functions properly. The issues noted are improvements rather than bugs blocking playability.

Code Quality Comparison

Aspect 27B Q4 35B-A3B Q3 35B-A3B Q4
Class structure Good Good Good
All 7 pieces Yes Yes Yes
Rotation states 4 each 4 each 4 each
Line clearing Yes Yes Yes
Scoring Yes Yes Yes
Game over Yes Yes Yes
Controls help Yes Yes Yes

All three models produced structurally similar, fully-featured implementations.

Play the Games

# 27B Q4 version
python docs/benchmarks/20260225_qwen35_comparison/tetris_qwen35-27b-multi.py

# 35B-A3B Q3 version (recommended - fastest model)
python docs/benchmarks/20260225_qwen35_comparison/tetris_qwen35-35b-q3-multi.py

# 35B-A3B Q4 version
python docs/benchmarks/20260225_qwen35_comparison/tetris_qwen35-35b-multi.py

Controls: ← → move, ↑ rotate, ↓ soft drop, Space hard drop, q quit

Recommendation

Use Qwen3.5-35B-A3B Q3_K_XL as the daily driver.

  • 3.8x faster than Qwen3.5-27B
  • Uses less VRAM (16GB vs 17GB)
  • Produces equivalent quality code
  • Best VRAM efficiency of all tested models

The 27B model should be avoided — it's dramatically slower with no quality benefit.

Reproducing

# Sync presets
uv run gpumod preset sync

# Test each model (one at a time due to VRAM)
uv run gpumod service start qwen35-27b-multi
uv run python docs/benchmarks/20260225_qwen35_comparison/benchmark_tetris.py \
  --model qwen35-27b-multi --output docs/benchmarks/20260225_qwen35_comparison/

uv run gpumod service stop qwen35-27b-multi
uv run gpumod service start qwen35-35b-q3-multi
uv run python docs/benchmarks/20260225_qwen35_comparison/benchmark_tetris.py \
  --model qwen35-35b-q3-multi --output docs/benchmarks/20260225_qwen35_comparison/

uv run gpumod service stop qwen35-35b-q3-multi
uv run gpumod service start qwen35-35b-multi
uv run python docs/benchmarks/20260225_qwen35_comparison/benchmark_tetris.py \
  --model qwen35-35b-multi --output docs/benchmarks/20260225_qwen35_comparison/

# Generate charts
uv run python docs/benchmarks/20260225_qwen35_comparison/generate_charts.py

Files

File Description
benchmark_tetris.py Benchmark script
generate_charts.py Chart generator
result_qwen35-27b-multi.json 27B results
result_qwen35-35b-q3-multi.json 35B-A3B Q3 results
result_qwen35-35b-multi.json 35B-A3B Q4 results
results_combined.json Combined comparison
tetris_qwen35-*.py Generated Tetris games
charts/ Generated comparison charts