Methodology

Memory fit

It fits when the budget plus 0.5 GiB spare is within usable memory: 100% of GPU VRAM, 75% of Mac and other unified memory. A “24 GB” device is planned as 24 GiB. These allowances are editable planning estimates, not vendor specifications.

• Weights: the exact byte sum of the complete Hugging Face file set (all shards) for that quant, pinned to a revision. Separate projectors and draft models are not included. BF16 is never relabelled FP16; mixed MLX conversions are flagged. Legacy models without verified files use parameters × bits per weight, labelled Estimated.
• KV cache: from config.json, per sequence, FP16/BF16 cache: 2 × full-attention layers × KV heads × head dim × 2 bytes per token, plus sliding-window layers up to their window. Hybrid recurrent/linear layers do not grow with context but hold a fixed state; when that state cannot be sized it is Unknown and the fit is labelled conditional. Compressed or backend-specific attention without an established formula is “Cache unknown”.
• Status: Fits (at the chosen context), Tight (only below it, down to 2K), Cache unknown (weights fit, cache not sized), Too big.
• Best file per device: the highest precision that fits, in the order Q8_0 → Q5_K_M → Q4_K_M for GPUs and MLX 8-bit → MLX 4-bit for Macs (GGUF is selectable on Macs).

Ranking

Primary

Descending live Artificial Analysis Intelligence Index, highest-scoring named evaluated mode per model (reasoning or non-reasoning). Provider-estimated scores stay visibly flagged.

Arena

Arena overall text ratings shown separately; tie-break only for a mapped non-proprietary entry. Broadly overlapping uncertainty means adjacent ranks are not proven differences.

No blending

No conversion or average between Arena Elo-like ratings and AA index.

Eligibility

Verified public release within 2026 as of cutoff; exact day unknown if only month/year is sourced. Superseded same-product checkpoints retained without duplicate rank.

Community

Dated Reddit/HF evidence changes practical recommendations and caveats, not the numerical index. Community X fetches blocked; no X consensus invented.

Rank among researched eligible open-weight candidates, not a guarantee of exhaustive global best. Models on hosted benchmarks and local quantizations can differ.

On a device page, models that fit are listed in this order; unranked 2026 models (no AA index) follow, newest first. Legacy (pre-2026), superseded and date-unverified models are hidden unless “Include legacy” is on, and never appear in picks.

Labels

• MeasuredA value printed by a public source for this model, quant label and device (same chip and GPU-core count for Macs); links to it. Not reproduced by LocalFit; build, context and settings may differ.
• EstimatedComputed by LocalFit; shown with “~”. See how we estimate.
• UnknownNo source and no defensible estimate. Never shown as zero.
• HF filesExact file sizes from a pinned Hugging Face repository.

Sources

• Models: publisher model cards, config.json and release announcements (97 current 2026 records; 82 with an exact public release day; month-only dates are shown as such).
• Quality: Artificial Analysis (observed 2026-10-17) and the Arena text dataset.
• Hardware: vendor specification pages only; FP16 is dense tensor throughput where published; AI TOPS are not FP16. Apple GPU FP16 and power are not published and stay unknown.
• Speeds: 4,373 reported observations from llama.cpp discussions, benchmark repositories, forums and vendor posts, each with source URL and quote.

Known gaps

• Exact release day unknown for 18 2026 records.
• Recurrent state, compressed attention and multimodal projectors can add memory the budget cannot size.
• Measured data is sparse for new model × quant × device combinations; most speed cells are estimates.
• A quant file existing does not prove your runtime supports the architecture.
• X/Twitter sentiment could not be collected; community signals are Hugging Face, GitHub and Reddit threads.