Three models, one machine, evidence
September 12, 2026 benchmark on RTX 5070 Ti Laptop: 12,227 MiB measured, Ollama, context 4096, one request at a time. No extrapolation to the desktop version, an RTX 5090, or a Mac.
The old multi-machine matrices were removed because there were no artifacts allowing them to be verified. This report publishes measurements from this pilot, with a warm-up followed by five repetitions per model; those from the AI TOP ATOM series have their own section below.
| Model | Quantization | Median tok/s | Min–max | FR Code | French documents |
|---|---|---|---|---|---|
| qwen3:8b | Q4_K_M | 59.71 | 59.45–59.96 | Success | Facts and format passed |
| hermes3:8b | Q4_0 | 71.97 | 71.79–72.15 | Failure | Successful facts, different keys |
| mistral-nemo:12b | Q4_0 | 47.44 | 47.37–47.56 | Success | Facts and format passed |
What this test allows us to say
Hermes3 generates faster in this test, but its function doesn't correctly remove duplicates and sort the values. Qwen3 and Mistral Nemo pass this exercise. Throughput therefore doesn't demonstrate quality. A coding task and a synthetic extraction task aren't enough to rank models for every use case; the Hermes3 output also hits the 256-token limit.
Reproduce and examine the data
- Protocol, parameters, and limitations
- Complete responses and GPU placements — raw JSON
- All repetitions — CSV
- Results, identifiers, and quantizations
- Hardware, versions, and initial state
- Reproducible script
- Complete folder, including inputs and outputs
Prompt processing and generation are separated in the raw data. The prefix cache can influence the former; the displayed throughput concerns generation only. It is neither an energy measurement nor the application’s total latency.
AI TOP ATOM series (GIGABYTE), October 2026
Our series on the AI TOP ATOM (GB10 NVIDIA chip, 128 GB), generously provided by GIGABYTE, consists of 7 guides published once a week since October 8, 2026. Each guide publishes its data here: every table as it appears in the guide, in CSV format (semicolon separator, UTF-8, legend in the first line), along with the method and model fingerprints. Series data, infographics, and photos: license CC BY 4.0, free to reuse with attribution to BestLLMfor.
- Measurement method and dated addenda
- SHA-256 hashes of the 27 model files in guide 1, compared with Hugging Face
- Guide 1: which LLMs actually run on 128 GB: Table 1 · Table 2 · Table 3 · Table 4 · Table 5 · Table 6
Three types of data
“Measured by QuelLLM” refers to an experiment associated with its artifacts. “External measurement” must identify the author and protocol. “Estimate” refers in particular to the heuristic speeds and memory budgets from the configurator; these values do not come from this new test.
Methodology · Choose based on my machine · Install local AI for free