The local-AI hardware index.
Aggregated numbers from real silicon, Llama, Mistral, DeepSeek, Qwen, SDXL, Whisper. Sortable, filterable, with $/tok·s⁻¹ and perf/watt so you can decide what to actually buy.
Where these numbers come from: These benchmarks are aggregated from llama.cpp community runs, vendor-published numbers, vLLM logs, and MLPerf submissions. MyAIHardware does not yet operate an in-house benchmark lab, that's coming Q3 2026. See per-row source notes and our full methodology.
Filter by silicon class, quantization, price, and VRAM to find your perfect setup.
Rankings show the largest available group with matching workload, quantization, context, batch-one settings and recorded runtime. Other configurations remain in the JSON and CSV exports.
No comparable multi-device results
The current selection lacks two devices with matching workload, quantization, context, batch and documented runtime. No speed winner is assigned. Inspect the source-attributed records.
Quick reference for every benchmark we run, including what it actually measures.
We separate hardware speed from model quality. These upstream snapshots show what the broader model-eval world thinks is strong right now, while our local benchmark data shows what actually runs well on your machine.
Snapshot generated
May 28, 2026
Weekly automated ingest with source-level freshness metadata.
Open-weight shortlist
6 local-friendly picks
Filtered to open text models with published parameter sizes and plausible local fit.
Reference sources
3 feeds
OpenEvals, LM Arena community Elo, and LiveBench model judgments.
Best open models to run locally
Directional quality signal for open-weight models. We bias this slice toward text models with known parameter sizes so builders can map quality to hardware fit.
#1
microsoft/Phi-3-medium-4k-instruct
14B · est. 8.4 GB Q4
#2
Qwen/Qwen2-72B
73B · est. 43.6 GB Q4
#3
microsoft/Phi-3.5-mini-instruct
3.8B · est. 2.3 GB Q4
#4
internlm/internlm2_5-7b-chat
7.7B · est. 4.6 GB Q4
#5
microsoft/Phi-3-mini-4k-instruct
3.8B · est. 2.3 GB Q4
Community preference ceiling
Useful for the broad popularity and preference frontier. This is mostly hosted frontier-model context, not a local-fit recommender.
#1
claude-opus-4-6-thinking
anthropic · 27.5K votes
#2
claude-opus-4-6
anthropic · 29.2K votes
#3
gemini-3.5-flash
google · 5.9K votes
#4
claude-opus-4-7-thinking
anthropic · 12.9K votes
#5
gemini-3.1-pro-preview
google · 34.2K votes
Recent task-judged performance
Another external frontier reference. We surface it so readers can separate 'runs fast locally' from 'is judged strong on recent tasks'.
#1
claude-3-7-sonnet-20250219-base
1 task buckets · 29 evals
#2
claude-3-opus-20240229
1 task buckets · 29 evals
#3
o1-2024-12-17-high
1 task buckets · 29 evals
#4
claude-3-7-sonnet-20250219-thinking-64k
1 task buckets · 29 evals
#5
gemini-2.0-flash-thinking-exp-1219
1 task buckets · 29 evals
Every record ships with the workload, model, quantization, runtime notes, source, and test date we could verify from the cited run. Because this is a blended dataset rather than one locked lab harness, compare results within the same workload and verification tier.
This release combines cited public runs from llama.cpp, MLPerf, vendor disclosures, and community logs. We normalize the metadata and clearly tag verification status rather than pretending every row came from one in-house harness.
Each workload page shows a representative prompt set and runtime command so you can reproduce the class of test. Exact prompts, runtimes, and harness settings still depend on the cited source for each record.
When a model doesn't fit in VRAM we explicitly note partial CPU offload. Pure-GPU runs and offload runs are not combined in the same rank.
Schema lives in src/data/benchmarkDatabase.ts on GitHub. Every record carries a sourceNote and testedAt date. v1.4 is the current cut, and the downloadable API is generated from the same enriched dataset the app renders.
Vendor-published numbers keep their sourceNote and verification label so you can apply your own discount. We do not silently blur vendor claims, community runs, and lab-verified records into one confidence tier.
$/1k tok is derived from MSRP and observed throughput using the public formula shown here. Perf/W is tokens/sec divided by TDP. No hidden weights, no proprietary composite score.
Running an exotic setup, Strix Halo, 4×3090 in a frame, M3 Ultra at 512GB unified? We want the data. Submissions are reviewed and credited. Repeat contributors get early access to upcoming benchmark releases.
Weekly AI hardware news + deals
Cross-reference benchmarks with our hardware databases.