Gemma 4 31B on RTX 5090 workstation (1x 32 GB)
31.273B on 32 GB at 1,792 GB/s. Hardware details · Model details
Marginal. Gemma 4 31B fits on RTX 5090 workstation (1x 32 GB) at the recommended
Q4_K_M configuration, requiring approximately 27.2 GB at 8K context. Practical context capacity is 8K. Measured evidence: 67.31 t/s median across 33 comparable runs at Q4_K_M and 128 context. AggregatedCompatibilityBorderline
Calculated fit
Yes
Calculated Uses 90-97% of usable memory. Likely to load but fragile under load or long context.
Recommended quantization
Q4_K_M
Calculated gguf
Memory required
27.2 GB
Calculated of 29.4 GB usable — 93%
Max practical context
8K
Calculated model supports 256K
Observed resident fit: Q4_K_M with eng_llama_cpp at 128 context. This is measured evidence for those exact artifacts and settings; it does not change the standard memory calculation above.
Memory budget at 8K context
Model weights ⓘ
18.2 GB Calculated
KV cache ⓘ
7.5 GB Calculated
Runtime overhead ⓘ
1.5 GB Estimated
Total required
27.2 GB Calculated
Headroom ⓘ
2.2 GB Calculated
Only a borderline fit (93% of usable memory). Expect to run headless and keep the context short.
Discrete GPU: 32 GB of VRAM, of which we assume 92% is usable after driver and context overhead.
Borderline fit. It may load, but you will have little room for a longer context, a second process, or a desktop session.
Performance
Measured performance · recommended Q4_K_M
Decode, prefill and TTFT below are measured for Q4_K_M, the recommended configuration.
Measured decode
67.3 t/s
Aggregated median; observed 64.7–69.7 t/s
Measured prefill
—
Aggregated not reported
TTFT at 8K
0.19 s
Aggregated median time to first token
Hardware load reference
640 W
Measured machine-level load; not this model run
Measurement evidence
33 measured runsRows are kept separate when quantization, runtime, context, workload or execution settings differ.
| Quant | Engine | Decode mode | Context | Runs | Median decode | Median prefill | Latest | Sources |
|---|---|---|---|---|---|---|---|---|
Q4_K_M | llama.cpp | standard | 128 | 33 | 67.31 t/s | — | 1 Jul 2026 | llm-speed.com |
How the estimate is calculated
- Decode: reading 31.3B active parameters at 4.85 bits/weight takes 13.36 ms at 1792 GB/s x 79% achieved efficiency.
- Prefill: 210 TFLOPS (FP16) x 2 for native FP8 tensor cores x 0.71 calibrated against measured prefill on this platform x 32% assumed model-FLOPs utilisation, divided by 2 x 31.3B parameters per token.
- Prefill throughput is highly engine-dependent. Flash attention, batch size and quantized KV all move this number substantially, and for sparse mixture-of-experts models it is the least reliable figure we produce.
- This is a calculated estimate, not a measurement. It assumes a single request, a short prompt, no speculative decoding and a warm model already resident in memory.
Estimator version estimator_v1. Stored with every estimated row so old estimates can be regenerated when the model improves.
Why this is rated "High" confidence
- Independent third-party benchmark: +3
- 33 measurements: +2
- Sources agree closely (±1%): +2
- Incomplete run metadata: +0
Score 7/10. Aggregated with the median, not the mean, so one outlier cannot move the headline figure.
This model at other quantizations on this machine
| Quantization | Resident | Disk | Total needed | Utilisation | Fit | Max context |
|---|---|---|---|---|---|---|
Q4_K_Mrecommended | 18.2 GB | 18.2 GB | 27.2 GB | 93% | BorderlineResident run observed | 8K |
Q5_K_M | 21.3 GB | 21.3 GB | 30.5 GB | 104% | Does not fit | 4K |
Q8_0 | 31.6 GB | 31.6 GB | 41.0 GB | 139% | Does not fit | 0 |
BF16 | 58.3 GB | 58.3 GB | 68.5 GB | 233% | Does not fit | 0 |
Cost of running Gemma 4 31B on this machineUnited States (federal) · C corporation · 8h/day
Monthly economic cost
$46
Calculated after tax
Codex/Claude Code
$100/mo
≈ $100/mo · local is $54 less
Net cash at purchase
$3,332
Calculated VAT not reclaimable
Total over 5 years
$2,731
Calculated after tax, after resale
Cost per USD/1M tokens
$5.34
Calculated 8.5M tokens/month
The $100 comparison uses the Codex Pro 5x / Claude Max 5x plans. A fixed planning conversion is used for USD. This compares monthly spend only: subscriptions have usage limits, local hardware has different capabilities and constraints, and taxes or regional pricing may change the charged amount. Prices checked 7 September 2026.
Purchase-price input: $3,332 in United States (federal), tax/VAT excluded · estimated · NL retail (system build) (converted catalogue estimate). This is the localized purchase input; the after-tax economic cost is calculated separately below.
Monthly breakdown
| Depreciation 2,832 over 5 years, straight-line to a 500 residual | $47.21 |
| Electricity 32.4 kWh/month at 0.140/kWh | $4.53 |
| Cost of capital 4.0%/yr on 1,916 average capital employed | $6.39 |
| Monthly cost before tax | $58.13 |
| Electricity tax shield Running costs are deductible business expenses | −$0.95 |
| First-year expensing §179 (100.0%) | −$11.66 |
| Monthly economic cost after tax | $45.51 |
Three different numbers, deliberately
Cash cost
$3,332
Money that leaves the bank account on day one, net of reclaimable VAT.
Accounting depreciation
$47.21/month
$2,832 written down over 5 years to a $500 residual.
After-tax economic cost
$45.51/month
Depreciation plus running costs plus cost of capital, less the tax those deductions save. This is the figure to compare between machines.
Assumptions and sources (verified 2026-09-07)
VAT rate
0% Official spec
VAT recoverable
0% Official spec
Federal C-corporation estimate at the flat 21% rate. Pass-through entities should use the sole-proprietor estimate as a rougher proxy.
Effective deduction rate
21.00% Calculated
Headline marginal rate 21.00%.
Depreciation
5 years, straight-line Assumption
Residual value
$500 (15%) Assumption
Two GPU generations later, the card is worth a fraction of its list price.
Electricity
$0.140/kWh Assumption
Average power draw
184 W Calculated
Load 640 W for 20% of powered hours, idle 70 W for the rest — a machine that is on is not generating tokens the whole time.
Investment allowances
§179 100.0% Official spec
Worth $700 in total — first-year expensing that replaces later tax depreciation.
Cost of capital
$6.39/month Assumption
- VAT rate: No US federal VAT · verified 2026-09-07
- Marginal tax rate: IRS Publication 542 — Corporations · verified 2026-09-07
- VAT recoverable fraction: site assumption · verified 2026-09-07
- Useful life: IRS Publication 946 — How To Depreciate Property · verified 2026-09-07
- Electricity price: Site assumption · verified 2026-09-07
Assumes the machine generates tokens 20% of its 176 powered hours per month. Cost per token scales inversely with this number — halve the utilisation and the cost per token doubles.
Federal planning estimate only, not tax advice. State and local income tax, sales/use tax and incentives are excluded. It assumes 100% business use, a Section 179 election, enough business income to use it, and that the full annual limit remains available.
Efficiency metrics
Decode per $1,000 spent
20.2 t/s
Calculated purchase price only
Decode per $100/month
147.9 t/s
Calculated after-tax ownership cost
Tokens per joule
0.37
Calculated same as tokens/s per watt
USD per 1M output tokens
$5.34
Calculated at 20% utilisation
Local versus hosted APIs
| Hosted model | USD/1M output | Break-even | API at your volume | Verdict | Comparison type |
|---|---|---|---|---|---|
| Qwen3.8 Flash Alibaba Cloud | $0.42 | 29.6M/mo | $13 | API cheaper | Different model |
| DeepSeek-V4.1 Flash DeepSeek | $0.60 | 25.3M/mo | $15 | API cheaper | Different model |
| DeepSeek-V4 Pro DeepSeek | $1.98 | 6.3M/mo | $62 | Local cheaper | Different model |
| DeepSeek V4 Pro (Together) Together AI | $4.40 | 2.1M/mo | $181 | Local cheaper | Different model |
| GLM-5.3 Z.ai | $4.40 | 2.9M/mo | $133 | Local cheaper | Different model |
| Claude Haiku 4.5 Anthropic | $5.00 | 3.5M/mo | $111 | Local cheaper | Different model |
Different-model comparison: the hosted model is not the model you would run locally. Treat this as a workload-quality trade-off, not a direct economic equivalence — the frontier model may complete a task in fewer tokens, or complete tasks the local model cannot. Break-even is the monthly output volume at which API spend equals the $46/month economic cost of owning this machine, assuming 8 input tokens per output token. Hosted prices are published in USD. This table is the canonical US default.
Gemma 4 31B on other hardware
| Hardware | Decode | Price | Fit |
|---|---|---|---|
| NVIDIA DGX H200 (8x H200, 1,128 GB) | ~313–451 t/s | $359,430 * | Comfortable |
| Quad RTX 5090 workstation (4x 32 GB) | ~68.1–98 t/s | $12,611 * | Comfortable |
| Lenovo ThinkStation PX (4x RTX PRO 6000, 384 GB) | ~59.4–85.5 t/s | $58,553 * | Comfortable |
| Dual RTX 5090 workstation (2x 32 GB) | ~45.2–65.1 t/s | $5,854 * | Comfortable |
| Dell Precision 7960 Rack (2x RTX PRO 6000, 192 GB) | ~41.4–59.6 t/s | $36,663 * | Comfortable |
| Dual used RTX 4090 workstation (48 GB) | ~38–54.6 t/s | $3,692 * | Comfortable |
Nearest alternatives to the RTX 5090 workstation (1x 32 GB)
Gemma 4 31B on Dual RTX 5090 workstation (2x 32 GB)64 GBGemma 4 31B on Quad RTX 5090 workstation (4x 32 GB)128 GBGemma 4 31B on Dual used RTX 4090 workstation (48 GB)48 GBGemma 4 31B on Used RTX 4090 workstation (24 GB)24 GBGemma 4 31B on RTX 3090 workstation (Ryzen 9 7950X, 64 GB)24 GB
Compare RTX 5090 workstation (1x 32 GB) against Dual RTX 5090 workstation (2x 32 GB) →
Compare RTX 5090 workstation (1x 32 GB) against Dual RTX 5090 workstation (2x 32 GB) →