The best model that actually fits your machine.

How this is computed

Memory fit. Weight memory is computed directly from the parameter count and the quantization's bits-per-weight: weights_GB = params_B × bpw / 8. We then add 20% headroom for the KV cache, activations and runtime overhead (roughly a 4K-context working set) and pick the highest-quality quantization that fits your stated memory. These are arithmetic estimates from published bits-per-weight, not measured throughput — actual usage varies with context length, batch size and runtime.

Task ranking. Models are ordered by the benchmark recorded on the artifact in our registry. Only models carrying a real recorded score for the selected task appear — nothing is inferred or filled in. Every score is visible on the signed benchmark leaderboard and traceable to its artifact.

Sources

• Quantization bits-per-weight — Hugging Face GGUF documentation, derived from the llama.cpp k-quants PR #1684 (Q4_K = 4.5 bpw, Q5_K = 5.5, Q6_K = 6.5625, Q2_K = 2.625, IQ4_XS = 4.25).
• Running GGUF models with Ollama from the Hub — Hugging Face × Ollama documentation.
• Parameter counts — resolved from each model repository's own published safetensors index via the Hugging Face API (safetensors.total), so the figure is the publisher's, not an estimate. Models whose weights aren't public keep no parameter count rather than a guess.
• Intelligence / Coding / Agentic indices, measured tokens-per-second and time-to-first-token — Artificial Analysis via their Data API (independent third-party evaluation, methodology published here).
• GPQA Diamond and AIME 2025 — recorded per-artifact in this registry and shown on the leaderboard.
• Why there is no Arena Elo ranking here. The registry's arena_score column currently mixes two incompatible scales — genuine LMArena Elo for some models and a 0–100 index for others. Ranking across them would put strong models below weak ones, so the picker omits it until the column is split by provenance and re-attested. It remains visible on individual artifact pages, labelled with its source.
• Device memory figures are the vendors' published capacities; where a device uses unified memory, the usable share for inference is lower than the total installed.

On throughput figures. Tokens/sec shown on a result is Artificial Analysis's median measured output rate on hosted endpoints — it tells you how fast that model generates in general, not what it will do on the hardware you selected. Local speed depends on your accelerator, quantization and runtime. We show the measurement and label it rather than extrapolating a number we cannot verify.

On local.ai. We evaluated local.ai as a per-device throughput source, but its data sits behind an invitation-only early-access gate and is not publicly retrievable, so nothing here is drawn from or attributed to it.