Gemma 4 26B-A4B on Radeon AI PRO R9700 workstation (32 GB)
25.806B (3.8B active) on 32 GB at 644 GB/s. Hardware details · Model details
CompatibilityComfortable
Fits
Yes
Calculated Uses under 80% of usable memory. Room for a long context and other work.
Recommended quantization
Q4_K_M
Calculated gguf
Memory required
18.3 GB
Calculated of 29.4 GB usable — 62%
Max practical context
32K
Calculated model supports 256K
Memory budget at 8K context
Model weights ⓘ
15.0 GB Calculated
KV cache ⓘ
1.9 GB Calculated
Runtime overhead ⓘ
1.5 GB Estimated
Total required
18.3 GB Calculated
Usable memory ⓘ
29.4 GB Assumption
Headroom
11.1 GB Calculated
Highest-precision quantization that leaves headroom: uses 62% of usable memory at 8K context.
Discrete GPU: 32 GB of VRAM, of which we assume 92% is usable after driver and context overhead.
Mixture of experts: all 25.806B parameters must be resident in memory even though only ~3.8B are active per token. Memory follows total parameters; speed follows active parameters.
PerformanceEstimated0/10
Decode (generation)
~157–226 t/s
Estimated calculated, not measured
Prefill (prompt)
~2480–5150 t/s
Estimated
TTFT at 8K
~1.7–3.4 s
Estimated time to first token
Power while generating
420 W
Measured 1.59 tokens/s per watt
These figures are estimated, not measured. Nobody has published a benchmark of Gemma 4 26B-A4B on this machine that we know of, so we calculate throughput from memory bandwidth and active parameter count. Estimates are always shown as a range. Expand the working below to see exactly how the number was produced.
How the estimate is calculated
- Decode: reading 3.8B active parameters at 4.85 bits/weight takes 3.77 ms at 644 GB/s x 95% achieved efficiency.
- MoE routing penalty of 15% applied: expert gathers are less bandwidth-efficient than a dense sweep.
- Prefill: 96 TFLOPS (FP16) x 2.46 calibrated against measured prefill on this platform x 32% assumed model-FLOPs utilisation, divided by 2 x 9.9B parameters per token.
- Prefill uses 9.9B effective parameters, not the 3.8B active in decode: a batch of hundreds of tokens routes across most of the expert pool.
- MoE decode depends on how well the engine batches expert gathers; real results vary more than for dense models.
- 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.
This model at other quantizations on this machine
| Quantization | Weights | Total needed | Utilisation | Fit | Max context |
|---|---|---|---|---|---|
MLX 4-bit | 14.1 GB | 17.4 GB | 59% | Comfortable | 32K |
Q4_K_Mrecommended | 15.0 GB | 18.3 GB | 62% | Comfortable | 32K |
Q8_0 | 26.0 GB | 29.7 GB | 101% | Does not fit | 0 |
BF16 | 48.1 GB | 52.4 GB | 178% | Does not fit | 0 |
Cost of running Gemma 4 26B-A4B on this machineUnited States (federal) · C corporation · 8h/day
Monthly economic cost
$34
Calculated after tax
Codex/Claude Code
$100/mo
≈ $100/mo · local is $66 less
Net cash at purchase
$2,521
Calculated VAT not reclaimable
Total over 5 years
$2,044
Calculated after tax, after resale
Cost per USD/1M tokens
$1.41
Calculated 24.2M 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.
Monthly breakdown
| Depreciation 2,143 over 5 years, straight-line to a 378 residual | $35.72 |
| Electricity 21.1 kWh/month at 0.140/kWh | $2.96 |
| Cost of capital 4.0%/yr on 1,450 average capital employed | $4.83 |
| Monthly cost before tax | $43.51 |
| Electricity tax shield Running costs are deductible business expenses | −$0.62 |
| First-year expensing §179 (100.0%) | −$8.82 |
| Monthly economic cost after tax | $34.06 |
Three different numbers, deliberately
Cash cost
$2,521
Money that leaves the bank account on day one, net of reclaimable VAT.
Accounting depreciation
$35.72/month
$2,143 written down over 5 years to a $378 residual.
After-tax economic cost
$34.06/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
$378 (15%) Assumption
Two GPU generations later, the card is worth a fraction of its list price.
Electricity
$0.140/kWh Assumption
Average power draw
120 W Calculated
Load 420 W for 20% of powered hours, idle 45 W for the rest — a machine that is on is not generating tokens the whole time.
Investment allowances
§179 100.0% Official spec
Worth $529 in total — first-year expensing that replaces later tax depreciation.
Cost of capital
$4.83/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
75.8 t/s
Calculated purchase price only
Decode per $100/month
561.3 t/s
Calculated after-tax ownership cost
Tokens per joule
1.59
Calculated same as tokens/s per watt
USD per 1M output tokens
$1.41
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 | 22.1M/mo | $37 | About equal | Different model |
| DeepSeek-V4 Flash DeepSeek | $0.66 | 14.1M/mo | $59 | Local cheaper | Different model |
| DeepSeek-V4 Pro DeepSeek | $1.98 | 4.7M/mo | $176 | Local cheaper | Different model |
| DeepSeek V4 Pro (Together) Together AI | $4.40 | 1.6M/mo | $514 | Local cheaper | Different model |
| GLM-5.3 Z.ai | $4.40 | 2.2M/mo | $378 | Local cheaper | Different model |
| Claude Haiku 4.5 Anthropic | $5.00 | 2.6M/mo | $315 | 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 $34/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 26B-A4B on other hardware
| Hardware | Decode | Price | Fit |
|---|---|---|---|
| RTX 5090 workstation (1x 32 GB) | ~303–435 t/s | €3.699 | Comfortable |
| Quad RTX 5090 workstation (4x 32 GB) | ~281–404 t/s | €13.999 | Comfortable |
| Dual RTX 5090 workstation (2x 32 GB) | ~216–311 t/s | €6.499 | Comfortable |
| Used RTX 4090 workstation (24 GB) | ~209–301 t/s | €2.299 | Comfortable |
| RTX PRO 6000 Blackwell workstation (96 GB) | ~188–270 t/s | €11.499 | Comfortable |
| RTX PRO 6000 Max-Q workstation (96 GB, 300 W) | ~188–270 t/s | €11.999 | Comfortable |
Nearest alternatives to the Radeon AI PRO R9700 workstation (32 GB)
Gemma 4 26B-A4B on RTX 5090 workstation (1x 32 GB)32 GBGemma 4 26B-A4B on Used RTX 4090 workstation (24 GB)24 GBGemma 4 26B-A4B on Dual used RTX 4090 workstation (48 GB)48 GBGemma 4 26B-A4B on Mac Studio M5 Max 36 GB36 GBGemma 4 26B-A4B on Dual RTX 5090 workstation (2x 32 GB)64 GB
Compare Radeon AI PRO R9700 workstation (32 GB) against RTX 5090 workstation (1x 32 GB) →
Compare Radeon AI PRO R9700 workstation (32 GB) against RTX 5090 workstation (1x 32 GB) →
Page quality score (why this page is or is not indexed)
10/13 — indexed. Generated pages are gated so we do not ask a search engine to rank a page with nothing computed to say. The directive is emitted in the page head via the metadata API, not in the body, so it is authoritative.
- Has a memory-fit calculation (+3)
- Has a performance estimate (+2)
- Has a price (+2)
- Has an ownership economics calculation (+2)
- Well connected (18 internal links) (+1)