RTX 5090 32GB

// weights AWQ 4-bit for Mixtral 8x7B (46.7B params)
weights = params × bits ÷ 8
        = 46.7 × 4.25 ÷ 8
        = 24.81 GB

// budget on RTX 5090 (32GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.80 GB    (runtime, cuda, allocator)
safety    = 4.80 GB    (15% of 32GB)
budget    = vram − safety − kv − overhead
          = 32 − 4.80 − 0.05 − 1.80
          = 25.35 GB

// fit decision
24.81 ≤ 25.35  → FITS
headroom  = 0.54 GB of weights budget left

// weights Q5_K_M for Qwen 3.6 35B A3B (35B params)
weights = params × bits ÷ 8
        = 35 × 5.5 ÷ 8
        = 24.06 GB

// budget on RTX 5090 (32GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.80 GB    (runtime, cuda, allocator)
safety    = 4.80 GB    (15% of 32GB)
budget    = vram − safety − kv − overhead
          = 32 − 4.80 − 0.05 − 1.80
          = 25.35 GB

// fit decision
24.06 ≤ 25.35  → FITS
headroom  = 1.29 GB of weights budget left

// weights Q5_K_M for Yi 34B (34B params)
weights = params × bits ÷ 8
        = 34 × 5.5 ÷ 8
        = 23.38 GB

// budget on RTX 5090 (32GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.80 GB    (runtime, cuda, allocator)
safety    = 4.80 GB    (15% of 32GB)
budget    = vram − safety − kv − overhead
          = 32 − 4.80 − 0.05 − 1.80
          = 25.35 GB

// fit decision
23.38 ≤ 25.35  → FITS
headroom  = 1.97 GB of weights budget left

// weights Q5_K_M for Qwen 2.5 32B (32.5B params)
weights = params × bits ÷ 8
        = 32.5 × 5.5 ÷ 8
        = 22.34 GB

// budget on RTX 5090 (32GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.80 GB    (runtime, cuda, allocator)
safety    = 4.80 GB    (15% of 32GB)
budget    = vram − safety − kv − overhead
          = 32 − 4.80 − 0.05 − 1.80
          = 25.35 GB

// fit decision
22.34 ≤ 25.35  → FITS
headroom  = 3.00 GB of weights budget left

// weights Q5_K_M for Qwen 2.5 Coder 32B (32.5B params)
weights = params × bits ÷ 8
        = 32.5 × 5.5 ÷ 8
        = 22.34 GB

// budget on RTX 5090 (32GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.80 GB    (runtime, cuda, allocator)
safety    = 4.80 GB    (15% of 32GB)
budget    = vram − safety − kv − overhead
          = 32 − 4.80 − 0.05 − 1.80
          = 25.35 GB

// fit decision
22.34 ≤ 25.35  → FITS
headroom  = 3.00 GB of weights budget left

// weights Q5_K_M for Qwen3 30B A3B (30.5B params)
weights = params × bits ÷ 8
        = 30.5 × 5.5 ÷ 8
        = 20.97 GB

// budget on RTX 5090 (32GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.80 GB    (runtime, cuda, allocator)
safety    = 4.80 GB    (15% of 32GB)
budget    = vram − safety − kv − overhead
          = 32 − 4.80 − 0.05 − 1.80
          = 25.35 GB

// fit decision
20.97 ≤ 25.35  → FITS
headroom  = 4.38 GB of weights budget left

// weights Q6_K for Qwen 3.6 27B (27B params)
weights = params × bits ÷ 8
        = 27 × 6.56 ÷ 8
        = 22.14 GB

// budget on RTX 5090 (32GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.80 GB    (runtime, cuda, allocator)
safety    = 4.80 GB    (15% of 32GB)
budget    = vram − safety − kv − overhead
          = 32 − 4.80 − 0.05 − 1.80
          = 25.35 GB

// fit decision
22.14 ≤ 25.35  → FITS
headroom  = 3.21 GB of weights budget left

// weights Q6_K for Gemma 4 26B A4B (26B params)
weights = params × bits ÷ 8
        = 26 × 6.56 ÷ 8
        = 21.32 GB

// budget on RTX 5090 (32GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.80 GB    (runtime, cuda, allocator)
safety    = 4.80 GB    (15% of 32GB)
budget    = vram − safety − kv − overhead
          = 32 − 4.80 − 0.05 − 1.80
          = 25.35 GB

// fit decision
21.32 ≤ 25.35  → FITS
headroom  = 4.03 GB of weights budget left

// weights FP8/INT8 for Mistral Small 3 (24B params)
weights = params × bits ÷ 8
        = 24 × 8 ÷ 8
        = 24.00 GB

// budget on RTX 5090 (32GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.80 GB    (runtime, cuda, allocator)
safety    = 4.80 GB    (15% of 32GB)
budget    = vram − safety − kv − overhead
          = 32 − 4.80 − 0.05 − 1.80
          = 25.35 GB

// fit decision
24.00 ≤ 25.35  → FITS
headroom  = 1.35 GB of weights budget left

// weights Q8_0 for gpt-oss 20B (20.9B params)
weights = params × bits ÷ 8
        = 20.9 × 8.5 ÷ 8
        = 22.21 GB

// budget on RTX 5090 (32GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.80 GB    (runtime, cuda, allocator)
safety    = 4.80 GB    (15% of 32GB)
budget    = vram − safety − kv − overhead
          = 32 − 4.80 − 0.05 − 1.80
          = 25.35 GB

// fit decision
22.21 ≤ 25.35  → FITS
headroom  = 3.14 GB of weights budget left

// weights Q8_0 for StarCoder2 15B (15B params)
weights = params × bits ÷ 8
        = 15 × 8.5 ÷ 8
        = 15.94 GB

// budget on RTX 5090 (32GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.80 GB    (runtime, cuda, allocator)
safety    = 4.80 GB    (15% of 32GB)
budget    = vram − safety − kv − overhead
          = 32 − 4.80 − 0.05 − 1.80
          = 25.35 GB

// fit decision
15.94 ≤ 25.35  → FITS
headroom  = 9.41 GB of weights budget left

// weights Q8_0 for Phi-4 (14.7B params)
weights = params × bits ÷ 8
        = 14.7 × 8.5 ÷ 8
        = 15.62 GB

// budget on RTX 5090 (32GB) at ctx 8K, conc 1, 15% safety
kv_cache  = 0.05 GB    (1× at ctx 8K)
overhead  = 1.80 GB    (runtime, cuda, allocator)
safety    = 4.80 GB    (15% of 32GB)
budget    = vram − safety − kv − overhead
          = 32 − 4.80 − 0.05 − 1.80
          = 25.35 GB

// fit decision
15.62 ≤ 25.35  → FITS
headroom  = 9.73 GB of weights budget left