{ "schema_version": "1.0", "site": "https://llmroofline.dev", "repository": "https://github.com/FrancescoPacifico/llm-roofline-db", "license": "CC-BY-4.0", "models_count": 32, "hardware_count": 28, "quantizations": [ { "id": "FP16", "label": "FP16 / BF16", "bpw": 16, "quality_tier": "Lossless Reference", "ppl_delta": "+0.000", "default_backend": "vllm / transformers", "notes": "Full 16-bit floating point baseline; compute-heavy and highest VRAM footprint." }, { "id": "Q8_0", "label": "Q8_0 (GGUF / 8-bit)", "bpw": 8.5, "quality_tier": "Near-Lossless", "ppl_delta": "+0.002", "default_backend": "llama.cpp / MLX", "notes": "Virtually indistinguishable from FP16; halves weight memory and doubles decode speed." }, { "id": "Q6_K", "label": "Q6_K (GGUF K-Quant)", "bpw": 6.57, "quality_tier": "High Precision", "ppl_delta": "+0.008", "default_backend": "llama.cpp", "notes": "Sweet spot for 24B-32B coding and reasoning models when VRAM allows." }, { "id": "Q5_K_M", "label": "Q5_K_M (GGUF K-Quant)", "bpw": 5.69, "quality_tier": "Balanced High", "ppl_delta": "+0.025", "default_backend": "llama.cpp", "notes": "Strong balance between reasoning retention and bandwidth savings." }, { "id": "Q4_K_M", "label": "Q4_K_M (GGUF Default)", "bpw": 4.85, "quality_tier": "Recommended 4-bit", "ppl_delta": "+0.080", "default_backend": "llama.cpp / Ollama", "notes": "Industry standard 4-bit K-quant with mixed Q6_K/Q5_K attention & output tensors." }, { "id": "MLX_4BIT", "label": "MLX 4-bit (Affine Group-64)", "bpw": 4.5, "quality_tier": "Apple Silicon Optimized", "ppl_delta": "+0.095", "default_backend": "mlx-lm", "notes": "Native Apple Metal 4-bit affine quantization with minimal block metadata overhead." }, { "id": "IQ4_XS", "label": "IQ4_XS (Importance-Weighted)", "bpw": 4.25, "quality_tier": "Compact 4-bit", "ppl_delta": "+0.130", "default_backend": "llama.cpp", "notes": "Importance-matrix quantized 4-bit format that trims ~12% weight memory vs Q4_K_M." }, { "id": "Q3_K_M", "label": "Q3_K_M (GGUF 3-bit)", "bpw": 3.91, "quality_tier": "Aggressive Compression", "ppl_delta": "+0.350", "default_backend": "llama.cpp", "notes": "Useful for fitting 70B+ models into 48GB systems; noticeable degradation on small dense models." } ], "models": [ { "slug": "qwen3-235b-a22b", "name": "Qwen3 235B-A22B", "family": "Qwen3", "hf_repo": "Qwen/Qwen3-235B-A22B", "openrouter_id": "qwen/qwen3-235b-a22b-2507", "release_date": "2025-04-29", "license": "Apache-2.0", "params_total_b": 235.1, "params_active_b": 22, "is_moe": true, "num_experts_total": 128, "num_experts_active": 8, "attention_type": "gqa", "num_layers": 94, "hidden_size": 4096, "num_q_heads": 64, "num_kv_heads": 4, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 40960, "api_prompt_usd_1m": 0.14, "api_completion_usd_1m": 0.8, "search_tier": 1, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 315044 }, { "slug": "qwen3-32b", "name": "Qwen3 32B", "family": "Qwen3", "hf_repo": "Qwen/Qwen3-32B", "openrouter_id": "qwen/qwen3-32b", "release_date": "2025-04-29", "license": "Apache-2.0", "params_total_b": 32.8, "params_active_b": 32.8, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 64, "hidden_size": 5120, "num_q_heads": 64, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 40960, "api_prompt_usd_1m": 0.08, "api_completion_usd_1m": 0.28, "search_tier": 1, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 3371612 }, { "slug": "qwen3-30b-a3b", "name": "Qwen3 30B-A3B", "family": "Qwen3", "hf_repo": "Qwen/Qwen3-30B-A3B", "openrouter_id": "qwen/qwen3-30b-a3b", "release_date": "2025-04-29", "license": "Apache-2.0", "params_total_b": 30.5, "params_active_b": 3.3, "is_moe": true, "num_experts_total": 128, "num_experts_active": 8, "attention_type": "gqa", "num_layers": 48, "hidden_size": 2048, "num_q_heads": 32, "num_kv_heads": 4, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 40960, "api_prompt_usd_1m": 0.12, "api_completion_usd_1m": 0.5, "search_tier": 1, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 1135065 }, { "slug": "qwen3-14b", "name": "Qwen3 14B", "family": "Qwen3", "hf_repo": "Qwen/Qwen3-14B", "openrouter_id": "qwen/qwen3-14b", "release_date": "2025-04-29", "license": "Apache-2.0", "params_total_b": 14.8, "params_active_b": 14.8, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 40, "hidden_size": 5120, "num_q_heads": 40, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 40960, "api_prompt_usd_1m": 0.12, "api_completion_usd_1m": 0.24, "search_tier": 1, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 13049351 }, { "slug": "qwen3-8b", "name": "Qwen3 8B", "family": "Qwen3", "hf_repo": "Qwen/Qwen3-8B", "openrouter_id": "qwen/qwen3-8b", "release_date": "2025-04-29", "license": "Apache-2.0", "params_total_b": 8.2, "params_active_b": 8.2, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 36, "hidden_size": 4096, "num_q_heads": 32, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 40960, "api_prompt_usd_1m": 0.04, "api_completion_usd_1m": 0.14, "search_tier": 2, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "seed_reference", "hf_downloads": 6050582 }, { "slug": "qwen2-5-coder-32b", "name": "Qwen2.5 Coder 32B", "family": "Qwen2.5", "hf_repo": "Qwen/Qwen2.5-Coder-32B-Instruct", "openrouter_id": "qwen/qwen-2.5-coder-32b-instruct", "release_date": "2024-11-12", "license": "Apache-2.0", "params_total_b": 32.8, "params_active_b": 32.8, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 64, "hidden_size": 5120, "num_q_heads": 40, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 32768, "api_prompt_usd_1m": 0.66, "api_completion_usd_1m": 1, "search_tier": 1, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 537003 }, { "slug": "qwen2-5-72b", "name": "Qwen2.5 72B Instruct", "family": "Qwen2.5", "hf_repo": "Qwen/Qwen2.5-72B-Instruct", "openrouter_id": "qwen/qwen-2.5-72b-instruct", "release_date": "2024-09-19", "license": "Qwen", "params_total_b": 72.7, "params_active_b": 72.7, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 80, "hidden_size": 8192, "num_q_heads": 64, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 32768, "api_prompt_usd_1m": 0.36, "api_completion_usd_1m": 0.4, "search_tier": 2, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 149093 }, { "slug": "qwen2-5-14b", "name": "Qwen2.5 14B Instruct", "family": "Qwen2.5", "hf_repo": "Qwen/Qwen2.5-14B-Instruct", "openrouter_id": "", "release_date": "2024-09-19", "license": "Apache-2.0", "params_total_b": 14.8, "params_active_b": 14.8, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 48, "hidden_size": 5120, "num_q_heads": 40, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 32768, "api_prompt_usd_1m": 0.06, "api_completion_usd_1m": 0.2, "search_tier": 2, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "seed_reference", "hf_downloads": 603839 }, { "slug": "qwen2-5-7b", "name": "Qwen2.5 7B Instruct", "family": "Qwen2.5", "hf_repo": "Qwen/Qwen2.5-7B-Instruct", "openrouter_id": "", "release_date": "2024-09-19", "license": "Apache-2.0", "params_total_b": 7.6, "params_active_b": 7.6, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 28, "hidden_size": 3584, "num_q_heads": 28, "num_kv_heads": 4, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 32768, "api_prompt_usd_1m": 0.04, "api_completion_usd_1m": 0.12, "search_tier": 2, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "seed_reference", "hf_downloads": 6517377 }, { "slug": "deepseek-r1-671b", "name": "DeepSeek-R1 671B", "family": "DeepSeek", "hf_repo": "deepseek-ai/DeepSeek-R1", "openrouter_id": "deepseek/deepseek-r1-0528", "release_date": "2025-01-20", "license": "MIT", "params_total_b": 671, "params_active_b": 37, "is_moe": true, "num_experts_total": 256, "num_experts_active": 8, "attention_type": "mla", "num_layers": 61, "hidden_size": 7168, "num_q_heads": 128, "num_kv_heads": 128, "head_dim": 128, "kv_lora_rank": 512, "qk_rope_head_dim": 64, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 163840, "api_prompt_usd_1m": 0.5, "api_completion_usd_1m": 2.15, "search_tier": 1, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 1649739 }, { "slug": "deepseek-v3-671b", "name": "DeepSeek-V3 671B", "family": "DeepSeek", "hf_repo": "deepseek-ai/DeepSeek-V3", "openrouter_id": "deepseek/deepseek-chat-v3-0324", "release_date": "2024-12-26", "license": "DeepSeek", "params_total_b": 671, "params_active_b": 37, "is_moe": true, "num_experts_total": 256, "num_experts_active": 8, "attention_type": "mla", "num_layers": 61, "hidden_size": 7168, "num_q_heads": 128, "num_kv_heads": 128, "head_dim": 128, "kv_lora_rank": 512, "qk_rope_head_dim": 64, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 163840, "api_prompt_usd_1m": 0.29, "api_completion_usd_1m": 1.14, "search_tier": 1, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 1409220 }, { "slug": "deepseek-r1-distill-70b", "name": "DeepSeek-R1 Distill Llama 70B", "family": "DeepSeek", "hf_repo": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B", "openrouter_id": "", "release_date": "2025-01-20", "license": "MIT", "params_total_b": 70.6, "params_active_b": 70.6, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 80, "hidden_size": 8192, "num_q_heads": 64, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 131072, "api_prompt_usd_1m": 0.2, "api_completion_usd_1m": 0.6, "search_tier": 1, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "seed_reference", "hf_downloads": 23953 }, { "slug": "deepseek-r1-distill-32b", "name": "DeepSeek-R1 Distill Qwen 32B", "family": "DeepSeek", "hf_repo": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B", "openrouter_id": "", "release_date": "2025-01-20", "license": "MIT", "params_total_b": 32.8, "params_active_b": 32.8, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 64, "hidden_size": 5120, "num_q_heads": 40, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 131072, "api_prompt_usd_1m": 0.1, "api_completion_usd_1m": 0.3, "search_tier": 1, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "seed_reference", "hf_downloads": 360664 }, { "slug": "deepseek-r1-distill-14b", "name": "DeepSeek-R1 Distill Qwen 14B", "family": "DeepSeek", "hf_repo": "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B", "openrouter_id": "", "release_date": "2025-01-20", "license": "MIT", "params_total_b": 14.8, "params_active_b": 14.8, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 48, "hidden_size": 5120, "num_q_heads": 40, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 131072, "api_prompt_usd_1m": 0.06, "api_completion_usd_1m": 0.2, "search_tier": 2, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "seed_reference", "hf_downloads": 133517 }, { "slug": "llama-4-maverick-400b-a17b", "name": "Llama 4 Maverick 400B-A17B", "family": "Llama 4", "hf_repo": "meta-llama/Llama-4-Maverick-17B-128E-Instruct", "hf_config_repo": "unsloth/Llama-4-Maverick-17B-128E-Instruct", "openrouter_id": "meta-llama/llama-4-maverick", "release_date": "2025-04-05", "license": "Llama-4", "params_total_b": 402, "params_active_b": 17, "is_moe": true, "num_experts_total": 128, "num_experts_active": 2, "attention_type": "gqa", "num_layers": 48, "hidden_size": 5120, "num_q_heads": 40, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 1048576, "api_prompt_usd_1m": 0.35, "api_completion_usd_1m": 1, "search_tier": 1, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 12667 }, { "slug": "llama-4-scout-109b-a17b", "name": "Llama 4 Scout 109B-A17B", "family": "Llama 4", "hf_repo": "meta-llama/Llama-4-Scout-17B-16E-Instruct", "hf_config_repo": "unsloth/Llama-4-Scout-17B-16E-Instruct", "openrouter_id": "meta-llama/llama-4-scout", "release_date": "2025-04-05", "license": "Llama-4", "params_total_b": 109, "params_active_b": 17, "is_moe": true, "num_experts_total": 16, "num_experts_active": 2, "attention_type": "gqa", "num_layers": 48, "hidden_size": 5120, "num_q_heads": 40, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 1048576, "api_prompt_usd_1m": 0.1, "api_completion_usd_1m": 0.3, "search_tier": 1, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 109443 }, { "slug": "llama-3-3-70b", "name": "Llama 3.3 70B Instruct", "family": "Llama 3", "hf_repo": "meta-llama/Llama-3.3-70B-Instruct", "hf_config_repo": "unsloth/Llama-3.3-70B-Instruct", "openrouter_id": "meta-llama/llama-3.3-70b-instruct", "release_date": "2024-12-06", "license": "Llama-3.3", "params_total_b": 70.6, "params_active_b": 70.6, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 80, "hidden_size": 8192, "num_q_heads": 64, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 131072, "api_prompt_usd_1m": 0.71, "api_completion_usd_1m": 0.71, "search_tier": 1, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 344904 }, { "slug": "llama-3-1-405b", "name": "Llama 3.1 405B Instruct", "family": "Llama 3", "hf_repo": "meta-llama/Llama-3.1-405B-Instruct", "openrouter_id": "", "release_date": "2024-07-23", "license": "Llama-3.1", "params_total_b": 405.9, "params_active_b": 405.9, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 126, "hidden_size": 16384, "num_q_heads": 128, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 131072, "api_prompt_usd_1m": 0.8, "api_completion_usd_1m": 1.2, "search_tier": 2, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "seed_reference", "hf_downloads": 7747 }, { "slug": "llama-3-1-70b", "name": "Llama 3.1 70B Instruct", "family": "Llama 3", "hf_repo": "meta-llama/Llama-3.1-70B-Instruct", "hf_config_repo": "NousResearch/Meta-Llama-3.1-70B-Instruct", "openrouter_id": "meta-llama/llama-3.1-70b-instruct", "release_date": "2024-07-23", "license": "Llama-3.1", "params_total_b": 70.6, "params_active_b": 70.6, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 80, "hidden_size": 8192, "num_q_heads": 64, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 131072, "api_prompt_usd_1m": 0.72, "api_completion_usd_1m": 0.72, "search_tier": 2, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 91879 }, { "slug": "llama-3-1-8b", "name": "Llama 3.1 8B Instruct", "family": "Llama 3", "hf_repo": "meta-llama/Llama-3.1-8B-Instruct", "hf_config_repo": "NousResearch/Meta-Llama-3.1-8B-Instruct", "openrouter_id": "meta-llama/llama-3.1-8b-instruct", "release_date": "2024-07-23", "license": "Llama-3.1", "params_total_b": 8.03, "params_active_b": 8.03, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 32, "hidden_size": 4096, "num_q_heads": 32, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 131072, "api_prompt_usd_1m": 0.02, "api_completion_usd_1m": 0.04, "search_tier": 1, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 4238569 }, { "slug": "gemma-3-27b", "name": "Gemma 3 27B IT", "family": "Gemma 3", "hf_repo": "google/gemma-3-27b-it", "hf_config_repo": "unsloth/gemma-3-27b-it", "openrouter_id": "google/gemma-3-27b-it", "release_date": "2025-03-12", "license": "Gemma", "params_total_b": 27.4, "params_active_b": 27.4, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 62, "hidden_size": 5376, "num_q_heads": 32, "num_kv_heads": 16, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 1024, "global_layer_ratio": 0.1667, "max_context": 131072, "api_prompt_usd_1m": 0.08, "api_completion_usd_1m": 0.35, "search_tier": 1, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 326029 }, { "slug": "gemma-3-12b", "name": "Gemma 3 12B IT", "family": "Gemma 3", "hf_repo": "google/gemma-3-12b-it", "hf_config_repo": "unsloth/gemma-3-12b-it", "openrouter_id": "google/gemma-3-12b-it", "release_date": "2025-03-12", "license": "Gemma", "params_total_b": 12.2, "params_active_b": 12.2, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 48, "hidden_size": 3840, "num_q_heads": 16, "num_kv_heads": 8, "head_dim": 256, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 1024, "global_layer_ratio": 0.1667, "max_context": 131072, "api_prompt_usd_1m": 0.05, "api_completion_usd_1m": 0.15, "search_tier": 2, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 403950 }, { "slug": "gemma-3-4b", "name": "Gemma 3 4B IT", "family": "Gemma 3", "hf_repo": "google/gemma-3-4b-it", "hf_config_repo": "unsloth/gemma-3-4b-it", "openrouter_id": "google/gemma-3-4b-it", "release_date": "2025-03-12", "license": "Gemma", "params_total_b": 4.3, "params_active_b": 4.3, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 34, "hidden_size": 2560, "num_q_heads": 8, "num_kv_heads": 4, "head_dim": 256, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 1024, "global_layer_ratio": 0.1667, "max_context": 131072, "api_prompt_usd_1m": 0.05, "api_completion_usd_1m": 0.1, "search_tier": 2, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 1010980 }, { "slug": "mistral-small-3-1-24b", "name": "Mistral Small 3.1 24B", "family": "Mistral", "hf_repo": "mistralai/Mistral-Small-3.1-24B-Instruct-2503", "openrouter_id": "mistralai/mistral-small-3.1-24b-instruct", "release_date": "2025-03-17", "license": "Apache-2.0", "params_total_b": 24, "params_active_b": 24, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 40, "hidden_size": 5120, "num_q_heads": 32, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 131072, "api_prompt_usd_1m": 0.351, "api_completion_usd_1m": 0.555, "search_tier": 1, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 141785 }, { "slug": "codestral-22b", "name": "Codestral 22B", "family": "Mistral", "hf_repo": "mistralai/Codestral-22B-v0.1", "openrouter_id": "", "release_date": "2024-05-29", "license": "MNPL-0.1", "params_total_b": 22.2, "params_active_b": 22.2, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 56, "hidden_size": 6144, "num_q_heads": 48, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 32768, "api_prompt_usd_1m": 0.1, "api_completion_usd_1m": 0.3, "search_tier": 2, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "seed_reference", "hf_downloads": 3888 }, { "slug": "mixtral-8x22b", "name": "Mixtral 8x22B Instruct", "family": "Mistral", "hf_repo": "mistralai/Mixtral-8x22B-Instruct-v0.1", "openrouter_id": "mistralai/mixtral-8x22b-instruct", "release_date": "2024-04-17", "license": "Apache-2.0", "params_total_b": 141, "params_active_b": 39, "is_moe": true, "num_experts_total": 8, "num_experts_active": 2, "attention_type": "gqa", "num_layers": 56, "hidden_size": 6144, "num_q_heads": 48, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 65536, "api_prompt_usd_1m": 2, "api_completion_usd_1m": 6, "search_tier": 2, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 36961 }, { "slug": "mixtral-8x7b", "name": "Mixtral 8x7B Instruct", "family": "Mistral", "hf_repo": "mistralai/Mixtral-8x7B-Instruct-v0.1", "openrouter_id": "", "release_date": "2023-12-11", "license": "Apache-2.0", "params_total_b": 46.7, "params_active_b": 12.9, "is_moe": true, "num_experts_total": 8, "num_experts_active": 2, "attention_type": "gqa", "num_layers": 32, "hidden_size": 4096, "num_q_heads": 32, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 32768, "api_prompt_usd_1m": 0.24, "api_completion_usd_1m": 0.24, "search_tier": 2, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "seed_reference", "hf_downloads": 325372 }, { "slug": "phi-4-14b", "name": "Phi-4 14B", "family": "Phi", "hf_repo": "microsoft/phi-4", "openrouter_id": "microsoft/phi-4", "release_date": "2024-12-12", "license": "MIT", "params_total_b": 14.7, "params_active_b": 14.7, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 40, "hidden_size": 5120, "num_q_heads": 40, "num_kv_heads": 10, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 16384, "api_prompt_usd_1m": 0.07, "api_completion_usd_1m": 0.14, "search_tier": 1, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 218593 }, { "slug": "glm-4-32b", "name": "GLM-4 32B 0414", "family": "GLM-4", "hf_repo": "zai-org/GLM-4-32B-0414", "openrouter_id": "", "release_date": "2025-04-14", "license": "MIT", "params_total_b": 32.3, "params_active_b": 32.3, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 61, "hidden_size": 6144, "num_q_heads": 48, "num_kv_heads": 2, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 32768, "api_prompt_usd_1m": 0.1, "api_completion_usd_1m": 0.3, "search_tier": 2, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "seed_reference", "hf_downloads": 8369 }, { "slug": "command-a-111b", "name": "Command A 111B", "family": "Cohere", "hf_repo": "CohereLabs/c4ai-command-a-03-2025", "hf_config_repo": "unsloth/c4ai-command-a-03-2025", "openrouter_id": "cohere/command-a", "release_date": "2025-03-13", "license": "CC-BY-NC-4.0", "params_total_b": 111, "params_active_b": 111, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 64, "hidden_size": 12288, "num_q_heads": 96, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 4096, "global_layer_ratio": 0.25, "max_context": 262144, "api_prompt_usd_1m": 2.5, "api_completion_usd_1m": 10, "search_tier": 2, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "openrouter_live", "hf_downloads": 1956 }, { "slug": "qwq-32b", "name": "QwQ 32B Reasoning", "family": "Qwen2.5", "hf_repo": "Qwen/QwQ-32B", "openrouter_id": "", "release_date": "2025-03-06", "license": "Apache-2.0", "params_total_b": 32.8, "params_active_b": 32.8, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 64, "hidden_size": 5120, "num_q_heads": 40, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 40960, "api_prompt_usd_1m": 0.1, "api_completion_usd_1m": 0.3, "search_tier": 1, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "seed_reference", "hf_downloads": 30334 }, { "slug": "nemotron-super-49b", "name": "Llama-3.3 Nemotron Super 49B", "family": "Nemotron", "hf_repo": "nvidia/Llama-3_3-Nemotron-Super-49B-v1", "openrouter_id": "", "release_date": "2025-03-18", "license": "NVIDIA-Open", "params_total_b": 49.9, "params_active_b": 49.9, "is_moe": false, "num_experts_total": 1, "num_experts_active": 1, "attention_type": "gqa", "num_layers": 49, "hidden_size": 8192, "num_q_heads": 64, "num_kv_heads": 8, "head_dim": 128, "kv_lora_rank": 0, "qk_rope_head_dim": 0, "sliding_window": 0, "global_layer_ratio": 1, "max_context": 131072, "api_prompt_usd_1m": 0.15, "api_completion_usd_1m": 0.4, "search_tier": 2, "verified_at": "2026-11-15T12:27:14Z", "api_price_source": "seed_reference", "hf_downloads": 35925 } ], "hardware": [ { "slug": "mac-mini-m4-16gb", "name": "Apple Mac Mini M4 (16 GB)", "vendor": "Apple", "category": "apple_silicon", "vram_gib": 16, "num_gpus": 1, "interconnect": "uma", "mem_bw_gbs": 120, "fp16_tflops": 8.6, "bw_efficiency": 0.72, "compute_efficiency": 0.56, "tdp_watts": 45, "idle_watts": 6, "street_price_usd": 599, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "amazon_mac_mini_m4", "bh_mac_mini_m4" ], "recommended_backend": "mlx", "notes": "Unified LPDDR5X-7500 memory; entry-level 8B Q4/MLX node with near-silent idle draw.", "usable_mem_gib": 10, "tuned_usable_mem_gib": 11.5 }, { "slug": "mac-mini-m4-24gb", "name": "Apple Mac Mini M4 (24 GB)", "vendor": "Apple", "category": "apple_silicon", "vram_gib": 24, "num_gpus": 1, "interconnect": "uma", "mem_bw_gbs": 120, "fp16_tflops": 8.6, "bw_efficiency": 0.72, "compute_efficiency": 0.56, "tdp_watts": 45, "idle_watts": 6, "street_price_usd": 799, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "amazon_mac_mini_m4_24", "bh_mac_mini_m4_24" ], "recommended_backend": "mlx", "notes": "Fits 14B Q6_K or 30B-A3B Q4 MoE locally; prefill is compute-limited at 8.6 TFLOPS.", "usable_mem_gib": 18, "tuned_usable_mem_gib": 19.5 }, { "slug": "mac-mini-m4-pro-48gb", "name": "Apple Mac Mini M4 Pro (48 GB)", "vendor": "Apple", "category": "apple_silicon", "vram_gib": 48, "num_gpus": 1, "interconnect": "uma", "mem_bw_gbs": 273, "fp16_tflops": 18.4, "bw_efficiency": 0.74, "compute_efficiency": 0.58, "tdp_watts": 95, "idle_watts": 8, "street_price_usd": 1799, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "amazon_mac_mini_m4_pro", "bh_mac_mini_m4_pro" ], "recommended_backend": "mlx", "notes": "273 GB/s LPDDR5X; strong sweet spot for 32B Q4/Q6 and 30B-A3B MoE on desktop.", "usable_mem_gib": 37.44, "tuned_usable_mem_gib": 42.24 }, { "slug": "mac-mini-m4-pro-64gb", "name": "Apple Mac Mini M4 Pro (64 GB)", "vendor": "Apple", "category": "apple_silicon", "vram_gib": 64, "num_gpus": 1, "interconnect": "uma", "mem_bw_gbs": 273, "fp16_tflops": 18.4, "bw_efficiency": 0.74, "compute_efficiency": 0.58, "tdp_watts": 95, "idle_watts": 8, "street_price_usd": 1999, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "amazon_mac_mini_m4_pro", "bh_mac_mini_m4_pro" ], "recommended_backend": "mlx", "notes": "64 GB UMA fits 70B Q4_K_M when wired limit is tuned via sysctl; ~5-6 tok/s on 70B dense.", "usable_mem_gib": 49.92, "tuned_usable_mem_gib": 56.32 }, { "slug": "macbook-pro-m3-max-64gb", "name": "Apple MacBook Pro M3 Max (64 GB)", "vendor": "Apple", "category": "apple_silicon", "vram_gib": 64, "num_gpus": 1, "interconnect": "uma", "mem_bw_gbs": 400, "fp16_tflops": 28.4, "bw_efficiency": 0.76, "compute_efficiency": 0.58, "tdp_watts": 115, "idle_watts": 12, "street_price_usd": 3499, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "amazon_mbp_m3_max", "bh_mac_studio_m4_max" ], "recommended_backend": "mlx", "notes": "40-core Metal GPU @ 400 GB/s; flagship portable workstation for 32B Q6_K and 70B IQ4_XS.", "usable_mem_gib": 49.92, "tuned_usable_mem_gib": 56.32 }, { "slug": "macbook-pro-m3-max-128gb", "name": "Apple MacBook Pro M3 Max (128 GB)", "vendor": "Apple", "category": "apple_silicon", "vram_gib": 128, "num_gpus": 1, "interconnect": "uma", "mem_bw_gbs": 400, "fp16_tflops": 28.4, "bw_efficiency": 0.76, "compute_efficiency": 0.58, "tdp_watts": 115, "idle_watts": 12, "street_price_usd": 4299, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "amazon_mbp_m3_max", "bh_mac_studio_m4_max" ], "recommended_backend": "mlx", "notes": "128 GB portable memory pool capable of running Llama 4 Scout 109B and Qwen3 235B IQ3/Q4.", "usable_mem_gib": 99.84, "tuned_usable_mem_gib": 112.64 }, { "slug": "mac-studio-m4-max-64gb", "name": "Apple Mac Studio M4 Max (64 GB)", "vendor": "Apple", "category": "apple_silicon", "vram_gib": 64, "num_gpus": 1, "interconnect": "uma", "mem_bw_gbs": 546, "fp16_tflops": 36.8, "bw_efficiency": 0.77, "compute_efficiency": 0.6, "tdp_watts": 140, "idle_watts": 10, "street_price_usd": 2499, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "bh_mac_studio_m4_max", "amazon_mac_studio" ], "recommended_backend": "mlx", "notes": "546 GB/s unified memory delivers ~18-20 tok/s on 32B Q4 and ~85 tok/s on Qwen3 30B-A3B.", "usable_mem_gib": 49.92, "tuned_usable_mem_gib": 56.32 }, { "slug": "mac-studio-m4-max-128gb", "name": "Apple Mac Studio M4 Max (128 GB)", "vendor": "Apple", "category": "apple_silicon", "vram_gib": 128, "num_gpus": 1, "interconnect": "uma", "mem_bw_gbs": 546, "fp16_tflops": 36.8, "bw_efficiency": 0.77, "compute_efficiency": 0.6, "tdp_watts": 140, "idle_watts": 10, "street_price_usd": 3499, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "bh_mac_studio_m4_max", "amazon_mac_studio" ], "recommended_backend": "mlx", "notes": "Best price-per-bandwidth 128 GB local workstation for 70B-111B models and 32k contexts.", "usable_mem_gib": 99.84, "tuned_usable_mem_gib": 112.64 }, { "slug": "mac-studio-m2-ultra-192gb", "name": "Apple Mac Studio M2 Ultra (192 GB)", "vendor": "Apple", "category": "apple_silicon", "vram_gib": 192, "num_gpus": 1, "interconnect": "uma", "mem_bw_gbs": 800, "fp16_tflops": 54.4, "bw_efficiency": 0.78, "compute_efficiency": 0.56, "tdp_watts": 215, "idle_watts": 14, "street_price_usd": 4999, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "bh_mac_studio_ultra", "amazon_mac_studio" ], "recommended_backend": "mlx", "notes": "800 GB/s dual-die Ultra interconnect; widely available used/refurbished for 120B-235B MoE.", "usable_mem_gib": 149.76, "tuned_usable_mem_gib": 168.96 }, { "slug": "mac-studio-m3-ultra-256gb", "name": "Apple Mac Studio M3 Ultra (256 GB)", "vendor": "Apple", "category": "apple_silicon", "vram_gib": 256, "num_gpus": 1, "interconnect": "uma", "mem_bw_gbs": 819, "fp16_tflops": 58, "bw_efficiency": 0.78, "compute_efficiency": 0.58, "tdp_watts": 230, "idle_watts": 15, "street_price_usd": 5599, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "bh_mac_studio_ultra", "amazon_mac_studio" ], "recommended_backend": "mlx", "notes": "Runs Qwen3 235B-A22B Q4_K_M + huge KV cache smoothly at ~35 tok/s under ~230W wall draw.", "usable_mem_gib": 199.68, "tuned_usable_mem_gib": 225.28 }, { "slug": "mac-studio-m3-ultra-512gb", "name": "Apple Mac Studio M3 Ultra (512 GB)", "vendor": "Apple", "category": "apple_silicon", "vram_gib": 512, "num_gpus": 1, "interconnect": "uma", "mem_bw_gbs": 819, "fp16_tflops": 58, "bw_efficiency": 0.78, "compute_efficiency": 0.58, "tdp_watts": 240, "idle_watts": 16, "street_price_usd": 9499, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "bh_mac_studio_ultra", "amazon_mac_studio" ], "recommended_backend": "mlx", "notes": "Fits DeepSeek-R1 671B Q4_K_M and Llama 4 Maverick 400B completely in single-box RAM.", "usable_mem_gib": 399.36, "tuned_usable_mem_gib": 450.56 }, { "slug": "ryzen-ai-max-395-128gb", "name": "Ryzen AI Max+ 395 Strix Halo (128 GB)", "vendor": "AMD / Beelink / GMKtec / Minisforum", "category": "uma_minipc", "vram_gib": 128, "num_gpus": 1, "interconnect": "uma", "mem_bw_gbs": 256, "fp16_tflops": 59.4, "bw_efficiency": 0.74, "compute_efficiency": 0.5, "tdp_watts": 140, "idle_watts": 14, "street_price_usd": 1899, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "gmktec_evo_x2", "beelink_gtr9_pro", "minisforum_ms_s1" ], "recommended_backend": "llama.cpp", "notes": "256-bit LPDDR5X-8000 UMA Mini-PC (GMKtec EVO-X2, Beelink GTR9 Pro, Minisforum MS-S1 Max). Up to 110 GiB VRAM via Linux TTM.", "usable_mem_gib": 96, "tuned_usable_mem_gib": 110 }, { "slug": "ryzen-ai-max-395-64gb", "name": "Ryzen AI Max+ 395 Strix Halo (64 GB)", "vendor": "AMD / Beelink / GMKtec", "category": "uma_minipc", "vram_gib": 64, "num_gpus": 1, "interconnect": "uma", "mem_bw_gbs": 256, "fp16_tflops": 59.4, "bw_efficiency": 0.74, "compute_efficiency": 0.5, "tdp_watts": 130, "idle_watts": 13, "street_price_usd": 1499, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "gmktec_evo_x2", "beelink_gtr9_pro" ], "recommended_backend": "llama.cpp", "notes": "Strong x86 ROCm/Vulkan alternative to Mac Mini M4 Pro at lower $/GB for 30B-A3B and 32B models.", "usable_mem_gib": 48, "tuned_usable_mem_gib": 54 }, { "slug": "ryzen-ai-hx-370-64gb", "name": "Ryzen AI 9 HX 370 Mini-PC (64 GB)", "vendor": "Beelink / Geekom / Minisforum", "category": "uma_minipc", "vram_gib": 64, "num_gpus": 1, "interconnect": "uma", "mem_bw_gbs": 120, "fp16_tflops": 23, "bw_efficiency": 0.7, "compute_efficiency": 0.45, "tdp_watts": 65, "idle_watts": 9, "street_price_usd": 899, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "beelink_ser9", "geekom_a9_max", "minisforum_ai" ], "recommended_backend": "llama.cpp", "notes": "Sub-$1k 64 GB x86 Mini-PC (Beelink SER9 / Geekom A9 Max); great for MoE A3B and 14B Q4.", "usable_mem_gib": 48, "tuned_usable_mem_gib": 54 }, { "slug": "rtx-5090-32gb", "name": "NVIDIA RTX 5090 (32 GB)", "vendor": "NVIDIA", "category": "discrete_gpu", "vram_gib": 32, "num_gpus": 1, "interconnect": "pcie", "mem_bw_gbs": 1792, "fp16_tflops": 209.5, "bw_efficiency": 0.82, "compute_efficiency": 0.65, "tdp_watts": 575, "idle_watts": 45, "street_price_usd": 2499, "hourly_price_usd": 0.89, "cloud_providers": [ "vast", "runpod" ], "affiliate_targets": [ "vast_gpu", "runpod_gpu", "amazon_rtx_5090", "bh_rtx_gpu" ], "recommended_backend": "vllm", "notes": "GDDR7 512-bit bus @ 1,792 GB/s; single-card speed king for 24B-32B models and massive prefill.", "usable_mem_gib": 31, "tuned_usable_mem_gib": 31 }, { "slug": "rtx-5080-16gb", "name": "NVIDIA RTX 5080 (16 GB)", "vendor": "NVIDIA", "category": "discrete_gpu", "vram_gib": 16, "num_gpus": 1, "interconnect": "pcie", "mem_bw_gbs": 960, "fp16_tflops": 112.6, "bw_efficiency": 0.82, "compute_efficiency": 0.64, "tdp_watts": 360, "idle_watts": 35, "street_price_usd": 1199, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "amazon_rtx_5080", "newegg_rtx", "bh_rtx_gpu" ], "recommended_backend": "llama.cpp", "notes": "Fast GDDR7 960 GB/s inference for 8B-14B Q4/Q6; 16 GiB ceiling restricts 24B+ models.", "usable_mem_gib": 15, "tuned_usable_mem_gib": 15 }, { "slug": "rtx-5070-ti-16gb", "name": "NVIDIA RTX 5070 Ti (16 GB)", "vendor": "NVIDIA", "category": "discrete_gpu", "vram_gib": 16, "num_gpus": 1, "interconnect": "pcie", "mem_bw_gbs": 896, "fp16_tflops": 87.9, "bw_efficiency": 0.82, "compute_efficiency": 0.64, "tdp_watts": 300, "idle_watts": 30, "street_price_usd": 849, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "amazon_rtx_5070_ti", "newegg_rtx", "bh_rtx_gpu" ], "recommended_backend": "llama.cpp", "notes": "High value 16 GB GDDR7 card (896 GB/s) outperforming RTX 4080 Super on decode speed.", "usable_mem_gib": 15, "tuned_usable_mem_gib": 15 }, { "slug": "rtx-5060-ti-16gb", "name": "NVIDIA RTX 5060 Ti (16 GB)", "vendor": "NVIDIA", "category": "discrete_gpu", "vram_gib": 16, "num_gpus": 1, "interconnect": "pcie", "mem_bw_gbs": 448, "fp16_tflops": 47.4, "bw_efficiency": 0.8, "compute_efficiency": 0.62, "tdp_watts": 180, "idle_watts": 22, "street_price_usd": 449, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "amazon_rtx_5060_ti", "newegg_rtx" ], "recommended_backend": "llama.cpp", "notes": "Most affordable 16 GB Blackwell card; great budget fit for 14B Q4_K_M coding assistants.", "usable_mem_gib": 15, "tuned_usable_mem_gib": 15 }, { "slug": "rtx-4090-24gb", "name": "NVIDIA RTX 4090 (24 GB)", "vendor": "NVIDIA", "category": "discrete_gpu", "vram_gib": 24, "num_gpus": 1, "interconnect": "pcie", "mem_bw_gbs": 1008, "fp16_tflops": 165.2, "bw_efficiency": 0.82, "compute_efficiency": 0.65, "tdp_watts": 450, "idle_watts": 35, "street_price_usd": 1999, "hourly_price_usd": 0.44, "cloud_providers": [ "vast", "runpod", "vultr" ], "affiliate_targets": [ "vast_gpu", "runpod_gpu", "novita_gpu", "amazon_rtx_4090" ], "recommended_backend": "vllm", "notes": "1,008 GB/s 24 GB workhorse; fits 32B Q4_K_M tight or 24B Q5_K_M with fast prefill.", "usable_mem_gib": 23, "tuned_usable_mem_gib": 23 }, { "slug": "rtx-3090-24gb", "name": "NVIDIA RTX 3090 (24 GB Used)", "vendor": "NVIDIA", "category": "discrete_gpu", "vram_gib": 24, "num_gpus": 1, "interconnect": "pcie", "mem_bw_gbs": 936, "fp16_tflops": 71, "bw_efficiency": 0.8, "compute_efficiency": 0.62, "tdp_watts": 350, "idle_watts": 35, "street_price_usd": 799, "hourly_price_usd": 0.22, "cloud_providers": [ "vast" ], "affiliate_targets": [ "vast_gpu", "amazon_rtx_3090" ], "recommended_backend": "llama.cpp", "notes": "Best $/tok/s ratio in the homelab used market (936 GB/s & 24 GB VRAM under $800).", "usable_mem_gib": 23, "tuned_usable_mem_gib": 23 }, { "slug": "dual-rtx-3090-48gb", "name": "2x NVIDIA RTX 3090 Rig (48 GB)", "vendor": "NVIDIA", "category": "discrete_gpu", "vram_gib": 24, "num_gpus": 2, "interconnect": "pcie", "mem_bw_gbs": 936, "fp16_tflops": 71, "bw_efficiency": 0.8, "compute_efficiency": 0.6, "tdp_watts": 680, "idle_watts": 65, "street_price_usd": 1950, "hourly_price_usd": 0.45, "cloud_providers": [ "vast" ], "affiliate_targets": [ "vast_gpu", "amazon_rtx_3090" ], "recommended_backend": "vllm", "notes": "Classic 48 GB homelab tensor-parallel rig for 70B Q4_K_M and 32B Q8_0 at high speed.", "usable_mem_gib": 46.65, "tuned_usable_mem_gib": 46.65 }, { "slug": "rtx-3060-12gb", "name": "NVIDIA RTX 3060 (12 GB)", "vendor": "NVIDIA", "category": "discrete_gpu", "vram_gib": 12, "num_gpus": 1, "interconnect": "pcie", "mem_bw_gbs": 360, "fp16_tflops": 25.5, "bw_efficiency": 0.78, "compute_efficiency": 0.6, "tdp_watts": 170, "idle_watts": 20, "street_price_usd": 289, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "amazon_rtx_3060", "newegg_rtx" ], "recommended_backend": "llama.cpp", "notes": "Most popular entry 12 GB GPU for 8B Q6/Q8 and 14B IQ4_XS tight context.", "usable_mem_gib": 11, "tuned_usable_mem_gib": 11 }, { "slug": "rx-7900-xtx-24gb", "name": "AMD Radeon RX 7900 XTX (24 GB)", "vendor": "AMD", "category": "discrete_gpu", "vram_gib": 24, "num_gpus": 1, "interconnect": "pcie", "mem_bw_gbs": 960, "fp16_tflops": 122.8, "bw_efficiency": 0.76, "compute_efficiency": 0.52, "tdp_watts": 355, "idle_watts": 35, "street_price_usd": 949, "hourly_price_usd": null, "cloud_providers": [], "affiliate_targets": [ "amazon_rx_7900_xtx", "newegg_amd" ], "recommended_backend": "llama.cpp", "notes": "960 GB/s 24 GB card with ROCm/Vulkan support; slower prefill kernels than CUDA Ada/Blackwell.", "usable_mem_gib": 23, "tuned_usable_mem_gib": 23 }, { "slug": "rtx-pro-6000-96gb", "name": "NVIDIA RTX PRO 6000 Blackwell (96 GB)", "vendor": "NVIDIA", "category": "discrete_gpu", "vram_gib": 96, "num_gpus": 1, "interconnect": "pcie", "mem_bw_gbs": 1792, "fp16_tflops": 250, "bw_efficiency": 0.84, "compute_efficiency": 0.68, "tdp_watts": 600, "idle_watts": 45, "street_price_usd": 8499, "hourly_price_usd": 1.85, "cloud_providers": [ "runpod", "vast" ], "affiliate_targets": [ "runpod_gpu", "vast_gpu", "bh_workstation_gpu" ], "recommended_backend": "vllm", "notes": "Single-slot/workstation 96 GB GDDR7 ECC powerhouse; fits 70B Q8_0 + 128k context in one card.", "usable_mem_gib": 95, "tuned_usable_mem_gib": 95 }, { "slug": "cloud-h200-sxm-141gb", "name": "Cloud NVIDIA H200 SXM (141 GB)", "vendor": "NVIDIA Cloud", "category": "cloud_gpu", "vram_gib": 141, "num_gpus": 1, "interconnect": "sxm", "mem_bw_gbs": 4800, "fp16_tflops": 989, "bw_efficiency": 0.84, "compute_efficiency": 0.72, "tdp_watts": 700, "idle_watts": 0, "street_price_usd": 36000, "hourly_price_usd": 3.49, "cloud_providers": [ "digitalocean", "runpod", "vast", "novita" ], "affiliate_targets": [ "digitalocean_gpu", "runpod_gpu", "vast_gpu", "novita_gpu" ], "recommended_backend": "vllm", "notes": "4.8 TB/s HBM3e single-GPU cloud node; effortlessly serves 70B FP16 or 235B Q4_K_M.", "usable_mem_gib": 140.4, "tuned_usable_mem_gib": 140.4 }, { "slug": "cloud-h100-sxm-80gb", "name": "Cloud NVIDIA H100 SXM (80 GB)", "vendor": "NVIDIA Cloud", "category": "cloud_gpu", "vram_gib": 80, "num_gpus": 1, "interconnect": "sxm", "mem_bw_gbs": 3350, "fp16_tflops": 989, "bw_efficiency": 0.84, "compute_efficiency": 0.72, "tdp_watts": 700, "idle_watts": 0, "street_price_usd": 28000, "hourly_price_usd": 2.39, "cloud_providers": [ "digitalocean", "runpod", "vast", "novita", "vultr" ], "affiliate_targets": [ "digitalocean_gpu", "runpod_gpu", "vast_gpu", "novita_gpu", "vultr_gpu" ], "recommended_backend": "vllm", "notes": "3.35 TB/s HBM3 cloud benchmark standard; ideal for high-batch vLLM serving & 70B Q4/Q6.", "usable_mem_gib": 79.4, "tuned_usable_mem_gib": 79.4 }, { "slug": "cloud-a100-sxm-80gb", "name": "Cloud NVIDIA A100 SXM (80 GB)", "vendor": "NVIDIA Cloud", "category": "cloud_gpu", "vram_gib": 80, "num_gpus": 1, "interconnect": "sxm", "mem_bw_gbs": 2039, "fp16_tflops": 312, "bw_efficiency": 0.82, "compute_efficiency": 0.68, "tdp_watts": 400, "idle_watts": 0, "street_price_usd": 14000, "hourly_price_usd": 1.39, "cloud_providers": [ "runpod", "vast", "novita", "vultr" ], "affiliate_targets": [ "vast_gpu", "runpod_gpu", "novita_gpu", "vultr_gpu" ], "recommended_backend": "vllm", "notes": "Value 80 GB HBM2e cloud workhorse for 70B Q4/Q8 and 109B Scout IQ4_XS.", "usable_mem_gib": 79.4, "tuned_usable_mem_gib": 79.4 }, { "slug": "cloud-l40s-48gb", "name": "Cloud NVIDIA L40S (48 GB)", "vendor": "NVIDIA Cloud", "category": "cloud_gpu", "vram_gib": 48, "num_gpus": 1, "interconnect": "pcie", "mem_bw_gbs": 864, "fp16_tflops": 362, "bw_efficiency": 0.82, "compute_efficiency": 0.66, "tdp_watts": 350, "idle_watts": 0, "street_price_usd": 8500, "hourly_price_usd": 0.89, "cloud_providers": [ "digitalocean", "runpod", "vast", "novita" ], "affiliate_targets": [ "digitalocean_gpu", "runpod_gpu", "vast_gpu", "novita_gpu" ], "recommended_backend": "vllm", "notes": "48 GB cloud inference sweet spot for 32B Q8_0 or 70B IQ4_XS at sub-$1/hr.", "usable_mem_gib": 47.4, "tuned_usable_mem_gib": 47.4 } ] }