AI Benchmarks

Head-to-head runs of self-hosted models on my own hardware.

Same model, different hardware, different number. The boards below are the best Elixir runs I have per machine: one fixed prompt, a needle test, and measured speed and code quality.

Leaderboard by hardware

Best Elixir runs per hardware.

The benchmark method

One task, one reply, six criteria, and as fair as possible across local model sizes.

The benchmarking prompt

It answers a single Elixir-specific question, so YMMV.

Top models by hardware

Full board

RX 9060 XT 16 GB

PCIe 3.0 x16

ModelQuantNeedleQuality
Ornith-1.5-35B-A3B MiniPlus V2.1 APEX-I MiniPlus V2.1 hit 2/5

RTX 4070 SUPER 12 GB + RTX 4060 Ti 16 GB

PCIe 4.0 x16 (CPU) + 4.0 x4 (PCH)

ModelQuantNeedleQuality
Groot-19B-A3B-261023 MTP Q5_K_M hit 2/5

2x RTX 5060 Ti 16 GB

PCIe 5.0 x8/x8

ModelQuantNeedleQuality
Nemotron-3.5-Bolt-74B-A9B IQ2 QAT IQ2 QAT hit 4/5
Muse-Glimmer NVFP4 NVFP4 hit 4/5
Ornith 1.5 35B-A3B NVFP4 NVFP4 hit 4/5
flint NVFP4 NVFP4 cliff 2/5
LFM2.5-24B-A2B Q8_0 Q8_0 hit 2/5

All runs

Date Model Hardware Quant Needle Quality
1 Nov 2026 Ornith-1.5-35B-A3B MiniPlus V2.1 RX 9060 XT 16 GB (single card) APEX-I MiniPlus V2.1 hit 2/5
31 Oct 2026 Groot-19B-A3B-261023 MTP RTX 4070 SUPER 12 GB + RTX 4060 Ti 16 GB (layer-split) Q5_K_M hit 2/5
31 Oct 2026 Nemotron-3.5-Bolt-74B-A9B IQ2 QAT 2× RTX 5060 Ti 16 GB (layer-split) IQ2 QAT hit 4/5
31 Oct 2026 Muse-Glimmer NVFP4 2× RTX 5060 Ti 16 GB (layer-split) NVFP4 hit 4/5
31 Oct 2026 flint NVFP4 2x RTX 5060 Ti 16 GB (layer-split) NVFP4 cliff 2/5
31 Oct 2026 LFM2.5-24B-A2B Q8_0 2x RTX 5060 Ti 16 GB (tensor-split) Q8_0 hit 2/5
30 Oct 2026 Ornith 1.5 35B-A3B NVFP4 2x RTX 5060 Ti 16 GB (tensor-split) NVFP4 hit 4/5

Have something worth building?

A paragraph about what you are working on is enough to start.

Say hello