mayank-dubey-ai/l4-gpu-llm-benchmark-leaderboard

🚀 Local LLM Serving & Quality Benchmark Leaderboard (NVIDIA L4 24GB)

An exhaustive, reproducible benchmark study measuring real-world serving performance (TTFT, TPOT, throughput, peak VRAM, energy consumption, and cost) alongside rigorous task quality gates (HumanEval+, MMLU-Pro, BFCL v4 tool calling, and RULER needle retrieval) for open-weight LLMs on a single NVIDIA L4 24GB GPU.

---

📊 Executive Summary & Key Takeaways

1. ⚡ Best Throughput & Coding Workhorse: openai/gpt-oss-20b (MXFP4 MoE)

   • Delivers 125.6 tok/s aggregate throughput at 512 C4 with an industry-low energy cost of ₹1.27 / 1M output tokens (@ ₹8/kWh).
   • Leads the coding category with 90.9% HumanEval and 87.2% strict HumanEval+.

2. 🧠 Best Reasoning & Function Calling: sokada4/Qwen3.8-27B (GPTQ Int4)

   • Outperforms in reasoning with 82.0% on MMLU-Pro (100-q stratified sample).
   • Undisputed champion on structured tool orchestration with 94.0% on BFCL v4 using native XML tool parsing.

3. 🏎️ Fastest Long-Context Prefill: nvidia/Nemotron-3.5-Lightning-30B (NVFP4 MoE/Mamba-2)

   • Delivers sub-3-second prefill at 8K context (2.74s TTFT) and 30.9 tok/s aggregate throughput at 8K C4, outperforming quadratic attention models by >15x on prefill latency.
   • Achieves 100.0% retrieval recall on 8K RULER needles.

4. ⚠️ 24GB VRAM Limits:

   • Dense 30B+ models (RedHatAI/Muse-Glimmer-30B) reach physical memory saturation on a 24GB card (21.93 GiB VRAM peak), bottlenecking decode speed (~6–7 tok/s) and causing queue explosion at long contexts.

---

🏆 1. Master Quality & Accuracy Leaderboard

Model	Architecture	Quantization Format	HumanEval (pass@1)	HumanEval+ (Strict)	MMLU-Pro (Reasoning)	BFCL Tool Calling	RULER 8K Retrieval	Evaluation Status
openai/gpt-oss-20b	21B MoE (3.6B active)	MXFP4 (Marlin MoE)	90.9% (149/164)	87.2% (143/164)	72.0% (72/100)	32.0% (16/50)	100.0% (30/30)	✅ Validated
sokada4/Qwen3.8-27B	27B Dense Hybrid	GPTQ Int4 (g128)	87.8% (144/164)	86.0% (141/164)	82.0% (82/100)	94.0% (47/50)	100.0% (30/30)	✅ Validated
nvidia/Nemotron-3.5-Lightning-30B	30B MoE (3.3B active / Mamba-2)	ModelOpt NVFP4/W4A16	50.0% (82/164)	48.8% (80/164)	40.0% (40/100)	20.0% (10/50)*	100.0% (30/30)	✅ Validated
RedHatAI/Muse-Glimmer-30B	30B Dense Multimodal	compressed-tensors W4A16	—	—	—	—	—	⚠️ INVALID**

*Note on Nemotron: Scored 10/10 on irrelevance detection, 0/40 on tool emission without proprietary adapter.
**Note on Muse-Glimmer: Quality scores invalidated due to upstream architecture wrapper mismatch in vLLM generic fallback mode.

---

⚡ 2. Serving Performance, Latency & Cost Matrix (Single L4)

A. Short Input Workload (512 Input Tokens, Concurrency 4)

Model	Format	TTFT (p50)	TPOT (p50)	Decode Speed	Aggregate Tok/s	Energy Wh / 1M	Cost INR / 1M (@ ₹8/kWh)
openai/gpt-oss-20b	MXFP4	0.49 s	27.03 ms	37.0 tok/s	125.6 tok/s	159 Wh	₹1.27
nvidia/Nemotron-3.5-Lightning-30B	NVFP4	0.46 s	78.63 ms	12.7 tok/s	46.0 tok/s	309 Wh	₹2.47
sokada4/Qwen3.8-27B	GPTQ Int4	2.83 s	68.83 ms	14.5 tok/s	42.3 tok/s	473 Wh	₹3.79
RedHatAI/Muse-Glimmer-30B	GPTQ W4A16	1.14 s	162.31 ms	6.2 tok/s	18.7 tok/s	956 Wh	₹7.65

B. Standard Context Workload (2K Input Tokens, Concurrency 4)

Model	Format	TTFT (p50)	TPOT (p50)	Aggregate Tok/s	Energy Wh / 1M	Cost INR / 1M
openai/gpt-oss-20b	MXFP4	1.51 s	27.85 ms	98.5 tok/s	202 Wh	₹1.62
nvidia/Nemotron-3.5-Lightning-30B	NVFP4	1.29 s	81.25 ms	41.5 tok/s	355 Wh	₹2.84
sokada4/Qwen3.8-27B	GPTQ Int4	7.94 s	81.55 ms	25.6 tok/s	777 Wh	₹6.22
RedHatAI/Muse-Glimmer-30B	GPTQ W4A16	4.08 s	219.24 ms	12.2 tok/s	1,554 Wh	₹12.43

C. Long Context Workload (8K Input Tokens, Concurrency 4)

Model	Format	TTFT (p50)	Aggregate Tok/s	Energy Wh / 1M	Cost INR / 1M	Long-Context Prefill Advantage
openai/gpt-oss-20b	MXFP4	3.79 s	48.3 tok/s	413 Wh	₹3.30	High sparse MoE throughput
nvidia/Nemotron-3.5-Lightning-30B	NVFP4	2.74 s	30.9 tok/s	533 Wh	₹4.27	Fastest 8K TTFT (Mamba-2 linear attention)
sokada4/Qwen3.8-27B	GPTQ Int4	42.36 s	8.7 tok/s	2,283 Wh	₹18.26	Standard quadratic attention prefill
RedHatAI/Muse-Glimmer-30B	GPTQ W4A16	22.64 s	1.4 tok/s	13,756 Wh	₹110.05	Saturated on 24GB L4 (1.44s/tok decode)

---

🛠️ Benchmark Suite & Evaluation Methodology

• Hardware Rig: NVIDIA L4 GPU (23,034 MiB VRAM), Driver 580.173.02, CUDA runtime 13.0.
• Inference Engine: vLLM 0.27.1 on PyTorch 2.13.0+cu130.
• Serving Metric Definitions:
  • TTFT (Time to First Token): Time from request submission to the first streamed token.
  • TPOT (Time per Output Token): Mean inter-token decode latency after the first token.
  • Aggregate Output Tok/s: Total output tokens across all concurrent streams divided by batch wall time.
  • GPU Energy (Wh / 1M tokens): Continuous NVML power sampling across batch execution.
• Quality Benchmarks:
  • HumanEval & HumanEval+: 164 problems scored via official evalplus.evaluate against v0.1.10 test vectors.
  • MMLU-Pro: Stratified 100-question sample across STEM, business, law, and economics.
  • BFCL v4: 50 non-live function calling tasks across 5 categories with AST JSON validation.
  • RULER / Needle: 90 cases across single-key and multi-key retrieval at 8K and 16K depths.

---

📖 Citation

If you utilize these benchmarks or findings in your research or deployment decisions, please cite:

@misc{dubey2026l4llmbenchmarks,
  author = {Mayank Dubey},
  title = {Local LLM Serving and Quality Benchmark Leaderboard on NVIDIA L4 GPU},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/mayank-dubey-ai/l4-gpu-llm-benchmark-leaderboard}}
}