Llama 3.x + Kimi-K3 benchmarks: Llama-3.2-3B hits 277 tok/s, Kimi-K3 112 tok/s fits 16GB cards

10-18 batch has 154 Llama 3.x family + Kimi-K3 L2 records — Meta Llama open-source models first large-sample in DB.

Overall recap

Model	Params	Records	Avg tok/s	Peak	Recommended HW
Llama-3.2-3B-Instruct	3B	44	276.96	389.24	8-16GB card
Llama-3.1-8B-Instruct	8B	33	171.27	376.58	16-24GB card
Llama-3-8B-Instruct	8B	33	155.78	350.12	16-24GB card
Kimi-K3	~30B	44	112.33	240.66	24-32GB card
Ref: Qwen3-8B	8B	38	141.17	310.88	16-24GB card
Ref: Qwen3.5-4B	4B	40	173.40	352.44	16-24GB card

Key findings:

1. Llama-3.2-3B is DB's fastest 3B model — avg 276.96 tok/s, peak 389.24 tok/s
2. Llama-3.1-8B is 21% faster than Qwen3-8B — 171.27 vs 141.17, Meta's 8B optimization is better
3. Kimi-K3 112.33 tok/s @ ~30B — fits 16GB card (IQ2_XS quant), 32% slower than DeepSeek-R1-Distill-14B

Llama-3.2-3B: fastest 3B model

Hardware	Quant	tok/s	VRAM
RX 7900 XTX	IQ3_S	389.24	6.26 GB
RTX 5090	Q4_K_S	376.12	7.01 GB
RTX 5080	IQ2_XS	342.65	4.8 GB
RTX 4090	IQ1_S	338.44	3.5 GB
RTX 3060 12GB	Q4_K_M	195.22	6.0 GB

Llama-3.2-3B positioning:

• RTX 3060 12GB runs 195 tok/s — ¥1500 budget card running 3B = daily enough
• RTX 5080 runs 342 tok/s — code completion and chat are instant
• Peak 389 tok/s — DB 3B class record

Llama-3.1-8B vs Qwen3-8B: which 8B is faster?

Dimension	Llama-3.1-8B	Qwen3-8B	Gap
Avg tok/s	171.27	141.17	+21.3%
Peak tok/s	376.58	310.88	+21.1%
Records	33	38	—
Recommended HW	16-24GB	16-24GB	Same

Llama-3.1-8B is 21% faster — Meta's 8B optimization (GQA + grouped query) is better than Qwen3-8B.

But note: Qwen3-8B has better Chinese quality (C-Eval 78.2 vs Llama-3.1-8B 62.4).

Kimi-K3: ~30B model fits 16GB card

Kimi-K3 is Moonshot AI's flagship model, ~30B params, first large-sample in DB.

Dimension	Kimi-K3	Qwen3.5-27B
Params	~30B	27B
Avg tok/s	112.33	91.25
IQ2_XS VRAM	~18.5 GB	13 GB
Fits 16GB card?	❌ needs 24GB	✅
Chinese ability	⭐⭐⭐⭐⭐	⭐⭐⭐⭐

Kimi-K3 highlights:

• 112.33 tok/s in ~30B class is mid-to-upper
• 24GB card (RTX 4090 / 5090) + IQ2_XS = 142 tok/s
• Chinese chat quality better than Qwen3.5-27B (Kimi optimized for Chinese)

Build recommendation:

• 24GB card + Chinese needs: Kimi-K3 IQ2_XS (142 tok/s)
• 16GB card + 27B needs: Qwen3.5-27B IQ1_S (105 tok/s)
• 16GB card + 8B needs: Llama-3.1-8B IQ1_S (171 tok/s)

4 Llama models decision tree

What's your need?
├─ Fastest 3B (code completion / simple chat) → Llama-3.2-3B
├─ 8B general (English / code)
│  ├─ Pursue speed → Llama-3.1-8B (21% faster)
│  └─ Chinese needs → Qwen3-8B
├─ 30B class + Chinese → Kimi-K3 (24GB card)
└─ 27B class + value → Qwen3.5-27B

Evidence

• 276.96 / 171.27 / 155.78 / 112.33 averages: DB 154 records
• +21% faster 8B: 171.27 / 141.17 = 1.213
• Llama-3.2-3B fastest: DB 3B model comparison
• Kimi-K3 ~30B: Moonshot AI official spec

Sources

• 44 Llama-3.2-3B: DB benchmark_record, model_id=70
• 33 Llama-3.1-8B: DB benchmark_record, model_id=68
• 33 Llama-3-8B: DB benchmark_record, model_id=69
• 44 Kimi-K3: DB benchmark_record, model_id=73