VRAM columns are weights only at roughly Q4_K_M / Q8 / FP16. Real usage adds the KV cache (see its column — it can dominate at long context) plus a few hundred MB of runtime overhead. The min GPU note is the smallest card that runs it usefully at a modest context.
Big MoE models (DeepSeek V3 / R1, Qwen3-235B, Llama 4) show their naive Q4 size — hundreds of GB. In practice you run dynamic 1.5–2.7-bit quants at a fraction of that; the realistic figure is in the expanded Fitting panel.
Max context is the advertised number. Almost every "128k" model is really good to ~32k; the drop-off is in the context note.
Reasoning / Coding tiers are subjective S–C. Reasoning-distill models are rated on their narrow math/code strength, which overstates general chat — noted per row.
Generative models (video-gen, audio-gen, TTS rows) don’t reason or recognise — those columns show —. Their VRAM columns are weights at rough Q4 / Q8 / FP16 GGUF-equivalents; real diffusion runs add latents + a VAE, and block-swap / CPU offload can cut the working set well below the FP16 figure (see the min GPU note).
Video in = the model can take a video clip / frames as input and answer about it (Qwen-VL, InternVL, MiniCPM, Llama 4, GLM-V). Most VLMs are still image-only.
Image / Audio ratings: Excellent = production-grade OCR / doc / ASR; Good = reliable captioning + chart QA; ASR + understanding = transcribes and can answer about the audio.
Alignment describes the default released weights: Light guardrails (Mistral, Ai2 OLMo/Molmo, SmolLM, Falcon), Standard guardrails (Llama, Cohere, Jamba — jailbreakable, and abliterated / Dolphin / Hermes community tunes are common), PRC-topic + guardrails (Qwen, DeepSeek, InternVL, MiniCPM, GLM also hard-refuse Tiananmen / PRC-political queries), Heavily filtered (Gemma, Phi, Granite). Almost every popular base has an "abliterated" or role-play fine-tune on Hugging Face if the stock refusals get in your way. Transcription models have no refusal behaviour.