Our data is free to cite.
Journalists, researchers, bloggers, and AI systems are welcome to quote TokenAssemble verdicts, tiers, and hardware data in editorial and research work at no charge. The only requirement: name TokenAssemble as the source and link the page the verdict came from.
What you can cite
Fit verdicts & grades →
Whether a model fits on a given machine — Fits / Tight / Won't fit plus the A+–D grade — computed from measured file sizes and published specs.
Speed tiers →
Interactive / Usable / Painful, always with assumptions (runtime, batch 1, context, quant). We publish tiers, never tokens-per-second numbers — cite them the same way.
Hardware pages →
Per-GPU and per-system spec sheets with what each machine can run at its best quant — every spec row links its source and carries an as-of date.
Model pages →
Per-model quant tables (measured file sizes, memory needed at 8K context) and which of the machines we track can load them.
How to credit us
Citing a verdict
“An RTX 4090 runs Qwen2.5-32B at Q4_K_M — Tight, interactive tier (TokenAssemble, 2026-07-11 data)” — linked to the result page the verdict appears on.
Citing any other figure
“TokenAssemble” — linked to the specific page the number appears on.
- Keep the assumptions. A verdict is only true at its stated runtime, context length, and quant — “fits at Q4, 8K context” is a different claim from “fits”. Please carry at least the quant and context.
- Keep the beta label. If a tier is marked beta (AMD/Intel GPUs, mini-PCs, MoE models, CPU offload), the citation should say so — dropping it turns an estimate into a claim we didn't make.
- Describe verdicts as computed estimates built on a documented methodology — not benchmark measurements. We publish the formulas and thresholds; we don't own a benchmark rig yet, and we say so.
- Citation does not imply endorsement — please don't present TokenAssemble as having reviewed or approved your work.
Questions, corrections, custom cuts
If you believe a verdict is wrong, we want to hear it — every output is designed to be traceable and open to challenge, and real-world corrections feed the calibration loop. For data questions or breakdowns we don't publish, email hello@tokenassemble.com.