Configurator v1
Will it fit?
Choose a model, a quant and a context length. We work out the memory it needs and which hardware can run it. Every setting is saved in the URL, so you can share the result.
Quantization compresses weights; lower precision saves memory and can reduce quality. 4.85 bits per weight · 16.6 GB file (HF API)
tok/s means generated tokens per second; roughly 4 English characters per token.
Defaults: η 0.60 GPU / 0.45 unified until calibrated; $0.15/kWh and 4 h/day from the brief's break-even example.1
Gemma 3 27B IT, Q4_K_M, 8k context
Where the memory goes
Weights 16.55 GB + KV cache 1.11 GB + overhead 0.56 GB. The reference device has 23.5 GB usable, leaving 5.3 GB free.
Hardware that fits
Listed price order, including flagged sample prices. Speed and fit are formula estimates, not buying recommendations. ±25% is an uncalibrated working assumption, not a measured confidence interval. Select up to three devices to compare, or show the working for one.
Purchase scope is shown beside each price. Electricity uses assumed whole-computer load power for 4 h/day at $0.15/kWh, 30 days/month; idle electricity is excluded. Build pages include the full parts list and idle electricity. Tax, shipping, maintenance and resale are excluded. Matching a subscription cost does not match its models, tools or limits.
Device | Fit | tok/s | First token | Price, USD | Break-even vs $200 |
|---|---|---|---|---|---|
| GeForce RTX 5060 Ti 16 GB 16 GB · 448 GB/s | Needs offload | 7.1 | 1.7 s | $779.99Sourced Cards only; host excluded | 4.0 moSample |
| GeForce RTX 4060 Ti 16 GB 16 GB · 288 GB/s | Needs offload | 5.8 | 1.8 s | $844.37Estimate Cards only; host excluded | 4.3 moSample |
| Radeon RX 7900 XTX 24 GB 24 GB · 960 GB/s | Fits | 32.6 | 652 ms | $899Sample Cards only; host excluded | 4.7 moSample |
| GeForce RTX 3090 24 GB (used) 24 GB · 936 GB/s | Fits | 31.8 | 1.1 s | $1,200Sourced Cards only; host excluded | 6.2 moSample |
| CPU only · Ryzen 9 9950X + 128 GB DDR5 128 GB · 90 GB/s | Fits | 2.5 | 32.1 s | $1,400Sample Cost scope undocumented | —Sample |
| 2× RTX 3090 (used) 48 GB 48 GB · 936 GB/s | Fits | 31.8 | 1.1 s | $2,400Estimate Cards only; host excluded | 12.9 moSample |
W = GGUF file size (HF API) = 16.55 GB KV = 2 × 62 layers × 16 KV heads × 128 × 8192 ctx × 1 × 2 B (52 layers capped at 1024) = 1.11 GB O = min(1.5, 0.5 + ctx/131072) = 0.56 GB [EST] M = W + KV + O = 18.22 GB U = 24 − 0.5 = 23.50 GB usable → Fits (fits ≤ 21.15) tps = η 0.6 × 960 GB/s / (W×1.000 + KV = 17.66 GB) = 32.6 tok/s ±25% pp = 0.35 × 123 TFLOPS / (2 × 27.43e9) = 785 tok/s; TTFT = 512 / pp = 652 ms E = 400 W × 4 h × 30 / 1000 × $0.15 = $7.20/mo; break-even = $899 / ($200 − E) = 4.7 months
Your decision summary
Gemma 3 27B IT at Q4_K_M · Chat · 8k context · F16 KV cache · 1 concurrent user(s) · No price limit · all platforms.
Estimated memory fit does not validate runtime, operating system, drivers or model support. Check the model card and runtime requirements before buying. Comparisons retain your choices even when filters change.
1 Formulas from the concept brief §7.3: memory = weights + KV cache + overhead; fits when total ≤ 90% of usable memory; decode ≈ η × bandwidth ÷ bytes read per token; offload adds the CPU share's time; prefill ≈ η × FLOPS ÷ (2 × active params); break-even = hardware ÷ (plan − electricity). Plan prices are US monthly, Oct 2026.
Model architecture and GGUF sizes are verified against the Hugging Face API. Device specs and prices are a mix of sourced and sample values; see each hardware datasheet. Estimates are not measurements. Runs entirely in your browser.