RTX 3070 8GB

// weights Q4_K_M for Qwen 3.5 9B (9B params)
weights = params × bits ÷ 8
        = 9 × 4.5 ÷ 8
        = 5.06 GB

// budget on RTX 3070 (8GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.20 GB    (runtime, cuda, allocator)
safety    = 1.20 GB    (15% of 8GB)
budget    = vram − safety − kv − overhead
          = 8 − 1.20 − 0.05 − 1.20
          = 5.55 GB

// fit decision
5.06 ≤ 5.55  → FITS
headroom  = 0.49 GB of weights budget left

// weights Q4_K_M for Gemma 2 9B (9B params)
weights = params × bits ÷ 8
        = 9 × 4.5 ÷ 8
        = 5.06 GB

// budget on RTX 3070 (8GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.20 GB    (runtime, cuda, allocator)
safety    = 1.20 GB    (15% of 8GB)
budget    = vram − safety − kv − overhead
          = 8 − 1.20 − 0.05 − 1.20
          = 5.55 GB

// fit decision
5.06 ≤ 5.55  → FITS
headroom  = 0.49 GB of weights budget left

// weights Q4_K_M for Llama 3.1 8B (8B params)
weights = params × bits ÷ 8
        = 8 × 4.5 ÷ 8
        = 4.50 GB

// budget on RTX 3070 (8GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.20 GB    (runtime, cuda, allocator)
safety    = 1.20 GB    (15% of 8GB)
budget    = vram − safety − kv − overhead
          = 8 − 1.20 − 0.05 − 1.20
          = 5.55 GB

// fit decision
4.50 ≤ 5.55  → FITS
headroom  = 1.05 GB of weights budget left

// weights Q4_K_M for Granite 8B Code (8B params)
weights = params × bits ÷ 8
        = 8 × 4.5 ÷ 8
        = 4.50 GB

// budget on RTX 3070 (8GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.20 GB    (runtime, cuda, allocator)
safety    = 1.20 GB    (15% of 8GB)
budget    = vram − safety − kv − overhead
          = 8 − 1.20 − 0.05 − 1.20
          = 5.55 GB

// fit decision
4.50 ≤ 5.55  → FITS
headroom  = 1.05 GB of weights budget left

// weights Q5_K_M for Mistral 7B v0.3 (7.2B params)
weights = params × bits ÷ 8
        = 7.2 × 5.5 ÷ 8
        = 4.95 GB

// budget on RTX 3070 (8GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.20 GB    (runtime, cuda, allocator)
safety    = 1.20 GB    (15% of 8GB)
budget    = vram − safety − kv − overhead
          = 8 − 1.20 − 0.05 − 1.20
          = 5.55 GB

// fit decision
4.95 ≤ 5.55  → FITS
headroom  = 0.60 GB of weights budget left

// weights Q5_K_M for Qwen 2.5 7B (7B params)
weights = params × bits ÷ 8
        = 7 × 5.5 ÷ 8
        = 4.81 GB

// budget on RTX 3070 (8GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.20 GB    (runtime, cuda, allocator)
safety    = 1.20 GB    (15% of 8GB)
budget    = vram − safety − kv − overhead
          = 8 − 1.20 − 0.05 − 1.20
          = 5.55 GB

// fit decision
4.81 ≤ 5.55  → FITS
headroom  = 0.74 GB of weights budget left

// weights Q8_0 for Gemma 4 E4B (4B params)
weights = params × bits ÷ 8
        = 4 × 8.5 ÷ 8
        = 4.25 GB

// budget on RTX 3070 (8GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.20 GB    (runtime, cuda, allocator)
safety    = 1.20 GB    (15% of 8GB)
budget    = vram − safety − kv − overhead
          = 8 − 1.20 − 0.05 − 1.20
          = 5.55 GB

// fit decision
4.25 ≤ 5.55  → FITS
headroom  = 1.30 GB of weights budget left

// weights Q8_0 for Phi-4 Mini (3.8B params)
weights = params × bits ÷ 8
        = 3.8 × 8.5 ÷ 8
        = 4.04 GB

// budget on RTX 3070 (8GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.20 GB    (runtime, cuda, allocator)
safety    = 1.20 GB    (15% of 8GB)
budget    = vram − safety − kv − overhead
          = 8 − 1.20 − 0.05 − 1.20
          = 5.55 GB

// fit decision
4.04 ≤ 5.55  → FITS
headroom  = 1.51 GB of weights budget left

// weights Q8_0 for Llama 3.2 3B (3.21B params)
weights = params × bits ÷ 8
        = 3.21 × 8.5 ÷ 8
        = 3.41 GB

// budget on RTX 3070 (8GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.20 GB    (runtime, cuda, allocator)
safety    = 1.20 GB    (15% of 8GB)
budget    = vram − safety − kv − overhead
          = 8 − 1.20 − 0.05 − 1.20
          = 5.55 GB

// fit decision
3.41 ≤ 5.55  → FITS
headroom  = 2.14 GB of weights budget left

// weights FP16/BF16 for Llama 3.2 1B (1.23B params)
weights = params × bits ÷ 8
        = 1.23 × 16 ÷ 8
        = 2.46 GB

// budget on RTX 3070 (8GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.20 GB    (runtime, cuda, allocator)
safety    = 1.20 GB    (15% of 8GB)
budget    = vram − safety − kv − overhead
          = 8 − 1.20 − 0.05 − 1.20
          = 5.55 GB

// fit decision
2.46 ≤ 5.55  → FITS
headroom  = 3.09 GB of weights budget left

// weights Q4_K_M for Phi-4 (14.7B params)
weights = params × bits ÷ 8
        = 14.7 × 4.5 ÷ 8
        = 8.27 GB

// budget on RTX 3070 (8GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.20 GB    (runtime, cuda, allocator)
safety    = 1.20 GB    (15% of 8GB)
budget    = vram − safety − kv − overhead
          = 8 − 1.20 − 0.05 − 1.20
          = 5.55 GB

// fit decision
8.27 > 5.55  → OVER
overflow  = 2.72 GB over budget

// weights Q4_K_M for StarCoder2 15B (15B params)
weights = params × bits ÷ 8
        = 15 × 4.5 ÷ 8
        = 8.44 GB

// budget on RTX 3070 (8GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.20 GB    (runtime, cuda, allocator)
safety    = 1.20 GB    (15% of 8GB)
budget    = vram − safety − kv − overhead
          = 8 − 1.20 − 0.05 − 1.20
          = 5.55 GB

// fit decision
8.44 > 5.55  → OVER
overflow  = 2.89 GB over budget