Local AI Model Database (2026): VRAM, Speed & Specs for 30+ Models

How to read this database

• Params — total parameters for dense models. For Mixture-of-Experts (MoE) models we list total / active (e.g. 30.5B total / 3.3B active): total decides memory, active decides speed.
• Context — the maximum context window from the official model card. Several models hit their largest context only with extrapolation (YaRN), noted inline.
• Min VRAM (Q4) — approximate memory to hold the weights at ~Q4_K_M quantization, plus a small KV-cache allowance. Long context, larger batches, or higher quants need more. For exact numbers against your GPU, use the VRAM calculator.
• tok/s ceiling (3090 / 4090) — the most tokens per second a dense model can generate on that card: NVIDIA's published memory bandwidth (936 GB/s for the 3090, 1,008 GB/s for the 4090) divided by the model's file size in the Ollama library, because every weight is read once per token. Real generation lands below the ceiling and drops as context grows. MoE models get no figure (only the active experts are read per token), and neither do models that do not fit in 24 GB, since layers spilled to CPU slow things sharply.
• License — the redistribution license. “Apache 2.0” and “MIT” are the most permissive; community/terms licenses (Llama, Gemma, NVIDIA Open Model) allow commercial use with conditions; Codestral (MNPL) and Mistral Large (MRL) are non-commercial — read before shipping.

The local AI model database (30+ models)

Grouped roughly by family, from smallest to largest within each. Names link to our full per-model pages where available.

VRAM = weights at ~Q4_K_M + small KV-cache allowance; add headroom for long context. tok/s are bandwidth ceilings (vendor memory bandwidth ÷ Ollama file size), not measurements. MoE rows show total / active params: total drives memory, active drives speed.

GPU reference table — what each card runs

VRAM capacity is the hard wall for local models; memory bandwidth then sets how fast they run. Prices are approximate US street prices as of mid-2026 and move constantly — check current listings before buying. To match a specific card to a specific model, use our which-GPU-to-buy tool or the can-I-run-it checker.

Prices are approximate US street estimates (mid-2026) and fluctuate; bandwidth from manufacturer specs. “Runs comfortably” assumes Q4-Q5 quants fully on the GPU with room for context.

VRAM by model class (the quick rule)

The honest caveat for MoE models: a model like Qwen3-235B-A22B only activates 22B params per token (so it generates at roughly 22B-class speed), but you must still hold all 235B params in memory. That is why “runs fast” and “fits in your VRAM” are two different questions for MoE — check both. The quantization calculator shows how each quant level (Q4/Q5/Q6/Q8) changes the footprint.

Methodology & honest caveats

• Specs are primary-sourced. Parameter counts, context windows and licenses come from each model official card, technical report, or vendor blog (see Sources). Where a family has multiple sizes we list the ones people actually run locally.
• tok/s are ceilings, not benchmarks. Each figure is the card's published memory bandwidth divided by the model's Ollama file size (read Nov 2, 2026). Your numbers will land below it and vary with quant, context length, batch size, backend (llama.cpp vs vLLM vs Ollama), driver and prompt.
• VRAM is the Q4 weight footprint plus a small KV allowance. Real usage climbs with context length. A 14B model that is ~9.5 GB of weights can need 11-12 GB once you load a long prompt.
• MoE memory ≠ MoE speed. Total params decide whether it fits; active params decide how fast it generates. We show both for every MoE row.
• Licenses change and have conditions. “Commercial OK” here is a summary, not legal advice — read the actual license (especially Codestral MNPL and Mistral Large MRL, which are non-commercial).
• This page is updated as models ship. The local landscape moves monthly; we revise the table rather than spawn dated duplicates.

Go deeper

Sources

Every parameter count, context window and license below is taken from the model official documentation. tok/s ceilings and VRAM are derived as described in Methodology: bandwidth from NVIDIA's GA102 and Ada whitepapers, file sizes from the Ollama library.

• Meta — Llama 3.1 release (8B / 70B / 405B, 128K context)
• Meta — Llama 3.3 70B model card
• Qwen — Qwen3 Technical Report (sizes, context, Apache 2.0)
• Qwen — Qwen3-Coder blog (480B-A35B / 30B-A3B, 256K)
• Qwen — Qwen2.5-Coder GitHub
• DeepSeek — DeepSeek-V3 Technical Report (671B / 37B active)
• DeepSeek — DeepSeek-R1 GitHub (distills, MIT)
• Google — Gemma 3 model card (1B-27B, 128K, multimodal)
• Mistral — Mistral Small 3 / 3.2 (24B, Apache 2.0)
• Mistral — Mistral NeMo 12B (Apache 2.0, 128K)
• Microsoft — gpt-oss / Phi context per official cards; Phi-4 (14B, MIT)
• OpenAI — Introducing gpt-oss (120b/20b, Apache 2.0, MXFP4)
• OpenAI — gpt-oss-120b model card (active params, 128K)
• Moonshot — Kimi K2 model card (1T / 32B active, 256K)
• Z.ai — GLM-4.6 model card (357B / 32B active, 200K, MIT)
• NVIDIA — Llama Nemotron paper (Nano 8B / Super 49B, 128K)
• NVIDIA / BIZON — Best GPU for local LLM 2026 (pricing, VRAM/$)