| # | PLAYER↕ | 🏆 WCS↕ | 🥊 PIR↕ | ⭐ SHIFT↕ | 🛡 Honesty↕ | 🧠 Intel↕ | 🏅 Union↕ | ⚖ League↕ | 📐 Params↕ | ⚡ tok/s↕ | 💾 RAM↕ | 🪤 Trap↕ | 📊 Calib↕ | 🚫 Refuse↕ | 🔄 Fix↕ | 🧩 Logic↕ | 🔢 Math↕ | 💻 Code↕ | 🌍 Lang↕ | 📚 Know↕ | 🧬 Meta↕ |
|---|
Upper-left = best value (high performance, low resource). Dot size = PIR. 🥅 League One models in the upper-left are Giant Killers.
Best model from each league, compared on 5 SHIFT axes. Outer = better.
Who squeezes the most speed from each GB of RAM? Bigger slice = more efficient.
Red zone = Frontier giants. Colored dots = Smol challengers. Closer to red = closer to SOTA.
A 4B model using only 2GB RAM achieves higher SHIFT scores than most 8B models requiring 5.5GB. Doubling parameters ≠ doubling performance.
→ SHIFT gap: 0.1 points for 2.75× more RAM
Qwen3-1.7B (1.2GB) outscores three 7-14B models. Latest architecture + small size > old architecture + big size.
→ 1.7B beats 7B, 8B, and 14B models
World's first 5-axis benchmark for small language models (≤10B active params). SHIFT measures what matters for edge: not just intelligence, but honesty, speed, and efficiency.
Size — Model footprint
Honesty — Hallucination, calibration, refusal, self-correction
Intelligence — Reasoning, math, coding, 7 languages, metacognition
Fast — Tokens/sec, TTFA
Thrift — Peak VRAM/RAM
WCS = √(SHIFT × PIRnorm)
The official ranking metric. Geometric mean of quality (SHIFT) and efficiency (PIR). Both must be high to score well.
PIR = (I × H × F) ÷ (S × T) · PIRnorm = log₁₀(PIR) / log₁₀(max) × 100
Efficiency rating. Like boxing's P4P: how much punch per pound of hardware.
🥅 League One (<2GB) — Raspberry Pi
⚽ La Liga (2-4GB) — Smartphone
🏅 Premier League (4-8GB) — Laptop
🏆 Champions League (8-16GB) — PC
🇬🇧 EN · 🇰🇷 KO · 🇸🇦 AR · 🇧🇷 PT · 🇹🇷 TR · 🇧🇩 BN · 🇹🇭 TH
2.7B+ speakers. Sentiment, idioms, translation, culture.
Same 20 cross-benchmark questions given to frontier SOTA models. Direct comparison with Claude, GPT-5, etc. Scores are not publicly disclosed.