Model benchmarks
phi3.5:3.8b local LLM performance
As of September 2026, phi3.5:3.8b runs at up to 117.7 tok/s for local inference (best of 2 community benchmark runs across 2 GPUs).
Model size
3.8B
Peak speed
117.7 tok/s
Average speed
64.1 tok/s
Avg PP
1951.1 tok/s
Min memory
3.9 GB
Max context
8.192 tokens
Avg output / run
7.510 tokens
Avg runtime / run
8m 31s
Avg quality
39.2
Benchmark runs
2
GPUs tested
2
Quality by task
Average LLM-judged quality (0–100) with the run-to-run spread shown as a P5–P95 band, overall and for each benchmark task, across all 2 runs. The low and high columns show how much the judge’s score varies between runs, and need at least two runs to display.
| Task | P5 (low) | Avg | P95 (high) |
|---|---|---|---|
| Overall | 38.9 | 39.2 | 39.6 |
| Agent Workflow | 60.1 | 64.5 | 68.8 |
| Code Generation | 2.9 | 6.1 | 9.3 |
| Role Play & Narrative | 46.4 | 48.2 | 50.0 |
| Research & Analysis | 36.8 | 38.2 | 39.6 |
Performance by hardware and tool
Every hardware/tool/quantization combination phi3.5:3.8b has been benchmarked on, ranked by peak token generation speed. Last updated September 2026.
| Hardware | Tool | Quant | Best tok/s | Avg tok/s | Memory | Context | Quality | Runs |
|---|---|---|---|---|---|---|---|---|
| NVIDIA GeForce RTX 3060 Ti | Ollama | Q4_0 | 117.7 tok/s | 117.7 tok/s | 5.0 GB | 8.192 tokens | 38.8 | 1 |
| NVIDIA GeForce RTX 2050 | Ollama | Q4_0 | 10.6 tok/s | 10.6 tok/s | 3.9 GB | 8.192 tokens | 39.7 | 1 |
Benchmark runs
All 2 phi3.5:3.8b runs submitted so far — expand one for its hardware, quality breakdown and per-scenario detail.
Frequently asked questions
- Is phi3.5:3.8b good for coding?
- In our benchmarks, phi3.5:3.8b scores 6.1/100 for coding. It runs at about 64.1 tok/s, so if you want more speed, unsloth/Qwen3.8-27B-GGUF:IQ3_S is faster (~104.3 tok/s) and still scores well for coding (83.0/100).
- Is phi3.5:3.8b good for agentic (tool-using) tasks?
- In our benchmarks, phi3.5:3.8b scores 64.5/100 for agentic workflows. It runs at about 64.1 tok/s, so if you want more speed, Tiel-Coder-35B-A3B-MTP-UD-IQ3_XXS is faster (~102.1 tok/s) and still scores well for agentic workflows (88.7/100).
- How fast is phi3.5:3.8b for local inference?
- Across 2 community benchmark runs, phi3.5:3.8b reaches up to 117.7 tok/s and averages 64.1 tok/s, with the fastest results on NVIDIA GeForce RTX 3060 Ti.
- How much memory does phi3.5:3.8b need?
- The leanest observed configuration used about 3.9 GB of memory (quantizations tested: Q4_0).
- Which tools have been used to run phi3.5:3.8b?
- Benchmarks were submitted using Ollama. Results are community-contributed and updated as new runs arrive.