model_id stringclasses 7
values | model_display_name stringclasses 7
values | publisher stringclasses 7
values | quantization stringclasses 1
value | format stringclasses 1
value | hardware dict | runtime dict | benchmark dict | evaluation dict | session dict |
|---|---|---|---|---|---|---|---|---|---|
nvidia/NVIDIA-Nemotron-3-Nano-4B-GGUF | nemotron-3-nano-4b | nvidia | Q4_K_M | GGUF | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "Windows 11"
} | {
"engine": "LM Studio",
"context_length": 16384,
"gpu_offload": "max",
"temperature": 0.7,
"top_p": 0.9
} | {
"tok_per_sec": 80.72,
"ttft_seconds": 0.37,
"tokens_generated": 1121,
"vram_used_gb": 3.64,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": 4,
"hit_length_target": true,
"notes": "Fastest small model in test set. Clean technical structure with horizontal-rule sections, math notation, accurate routing/activation/efficiency coverage. Edge-ready positioning per NVIDIA model card."
} | {
"date": "2026-05-05T00:00:00",
"run_type": "warm"
} |
google/gemma-4-e4b-it | gemma-4-e4b | google | Q4_K_M | GGUF | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "Windows 11"
} | {
"engine": "LM Studio",
"context_length": 16384,
"gpu_offload": "max",
"temperature": 0.7,
"top_p": 0.9
} | {
"tok_per_sec": 68.49,
"ttft_seconds": 0.26,
"tokens_generated": 901,
"vram_used_gb": 5.96,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": 4,
"hit_length_target": true,
"notes": "Fastest TTFT in test set (0.26s). Concise and accurate, hit length target. VRAM higher than 4B name suggests — E variants use selective activation but full weights still occupy VRAM."
} | {
"date": "2026-05-05T00:00:00",
"run_type": "warm"
} |
meta-llama/Llama-3.3-8B-Instruct | llama-3.3-8b | meta | Q4_K_M | GGUF | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "Windows 11"
} | {
"engine": "LM Studio",
"context_length": 16384,
"gpu_offload": "max",
"temperature": 0.7,
"top_p": 0.9
} | {
"tok_per_sec": 52.54,
"ttft_seconds": 0.39,
"tokens_generated": 3769,
"vram_used_gb": 4.95,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": 2,
"hit_length_target": false,
"notes": "Overshot length budget by 7.5x (3769 tokens for 500-word brief). Repeated entire sections multiple times. Fabricated specific benchmarks unprompted: '7.9x faster inference', '5B parameter MoE 1.97s vs Dense 15.9s' — no source cited, numbers appear invented... | {
"date": "2026-05-05T00:00:00",
"run_type": "warm"
} |
ibm-granite/granite-4.1-8b | granite-4.1-8b | ibm | Q4_K_M | GGUF | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "Windows 11"
} | {
"engine": "LM Studio",
"context_length": 16384,
"gpu_offload": "max",
"temperature": 0.7,
"top_p": 0.9
} | {
"tok_per_sec": 49.12,
"ttft_seconds": 0.32,
"tokens_generated": 913,
"vram_used_gb": 5.3,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": 5,
"hit_length_target": true,
"notes": "Best instruction-follower in test set. Hit ~500-word target near-exactly (913 tokens including markdown). Cleanest technical exposition. IBM 'matches our previous 32B MoE' claim is credible from this sample. Practitioner pick for accuracy."
} | {
"date": "2026-05-05T00:00:00",
"run_type": "warm"
} |
Qwen/Qwen3.5-9B | qwen3.5-9b | qwen | Q4_K_M | GGUF | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "Windows 11"
} | {
"engine": "LM Studio",
"context_length": 16384,
"gpu_offload": "max",
"temperature": 0.7,
"top_p": 0.9
} | {
"tok_per_sec": 44.17,
"ttft_seconds": 0.46,
"tokens_generated": 3589,
"vram_used_gb": 6.19,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": 4,
"hit_length_target": false,
"notes": "Reference baseline. Long but coherent. Only model in test set that called out memory-bandwidth bottleneck on consumer GPUs specifically — rare insight at this size class. Multimodal-capable. Validated reproducibility: 44.17 t/s today vs 45.41 t/s yesterday... | {
"date": "2026-05-05T00:00:00",
"run_type": "warm"
} |
microsoft/phi-4-mini-instruct | phi-4-mini-instruct | microsoft | Q4_K_M | GGUF | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "Windows 11"
} | {
"engine": "LM Studio",
"context_length": 16384,
"gpu_offload": "max",
"temperature": 0.7,
"top_p": 0.9
} | {
"tok_per_sec": 88.92,
"ttft_seconds": 0.38,
"tokens_generated": 889,
"vram_used_gb": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": 4,
"hit_length_target": false,
"notes": "Coherent on-topic prose, slight markdown rendering quirks. Length budget overshoot (target 500w, generated ~775w). New leader in 8GB VRAM dense Q4 catalog at 88.92 tok/sec, beating prior #1 nemotron-3-nano-4b (80.72)."
} | {
"date": "2026-05-06T00:00:00",
"run_type": "warm"
} |
mistralai/ministral-3-8b-instruct-2512 | ministral-3-8b | mistralai | Q4_K_M | GGUF | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "Windows 11"
} | {
"engine": "LM Studio",
"context_length": 16384,
"gpu_offload": "max",
"temperature": 0.7,
"top_p": 0.9
} | {
"tok_per_sec": 48.47,
"ttft_seconds": 0.57,
"tokens_generated": 889,
"vram_used_gb": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": 4,
"hit_length_target": true,
"notes": "Well-structured prose with clean sectioning. Better instruction-following than llama-3.3-8b at similar tok/sec — 4x fewer tokens for the same answer = much better wall-clock-per-useful-output."
} | {
"date": "2026-05-06T00:00:00",
"run_type": "warm"
} |
Local LLM Bench — RTX 4060 Ti 8GB
Real practitioner benchmarks of open-source LLMs on consumer 8GB VRAM hardware.
Hardware
- GPU: NVIDIA GeForce RTX 4060 Ti (8GB VRAM)
- CPU: AMD Ryzen 5 7600X (6 cores, AM5)
- RAM: 32GB DDR5-6000 CL36
- Platform: Windows 11
- Runtime: LM Studio (CUDA backend)
Methodology
All models loaded with:
- Quantization: Q4_K_M (GGUF)
- Context length: 16384 tokens
- GPU offload: maximum (full GPU residency where it fits)
- Temperature: 0.7
- Top-p: 0.9
Single benchmark prompt (prompts/moe-explainer-500w.md):
Explain how mixture-of-experts (MoE) inference works in modern LLMs. Cover: routing mechanism, expert activation per token, why MoE is more efficient than dense models of equal parameter count, and practical tradeoffs for local deployment on consumer hardware. Aim for ~500 words.
For each model, captured:
- Tokens/sec (decode speed, warm)
- Time-to-first-token (TTFT)
- Tokens generated total
- VRAM used (model + KV cache)
- Quality score (1-5, subjective)
- Length-target adherence (hit / overshoot / undershoot)
- Notes on output character
Format
JSONL, one row per benchmark run. See data/bench-2026-05-05.jsonl.
Full model responses in responses/<model-slug>/2026-05-05.md.
Intended use
- Reproducibility: methodology and prompts documented
- Reference for 8GB VRAM hardware planners
- Community contribution: add your own rows via PR
Limitations
- Single prompt — quality scores reflect performance on one task type (technical explanation). Coding, multilingual, long-context not tested here.
- Subjective quality scores. See
responses/for raw outputs to judge yourself. - Llama-3.3-8B response contained fabricated benchmark numbers — flagged in notes; use with caution for factual tasks.
Citation
If you use this data, cite as:
Witcheer (2026). Local LLM Bench — RTX 4060 Ti 8GB. https://huggingface.co/datasets/witcheer/windows-rtx-4060ti-8gb-bench-2026-05
Contact
- Downloads last month
- 89