gpt-oss-20b on RTX 3090 workstation (Ryzen 9 7950X, 64 GB)

20.915B (3.6B active) on 24 GB at 936 GB/s. Hardware details · Model details

Run locally →
Yes. gpt-oss-20b fits on RTX 3090 workstation (Ryzen 9 7950X, 64 GB) at the recommended MXFP4 configuration, requiring approximately 13.0 GB at 8K context. Practical context capacity is 128K. Expected decode for the recommended configuration is ~205–295 t/s. Estimated
CompatibilityComfortable
Calculated fit
Yes
Calculated Uses under 80% of usable memory. Room for a long context and other work.
Recommended quantization
MXFP4
Calculated gguf
Memory required
13.0 GB
Calculated of 22.1 GB usable — 59%
Max practical context
128K
Calculated model supports 128K

Memory budget at 8K context

Model weights
11.3 GB Calculated
KV cache
0.4 GB Calculated
Runtime overhead
1.3 GB Estimated
Total required
13.0 GB Calculated
Headroom
9.1 GB Calculated
Highest-precision quantization that leaves headroom: uses 59% of usable memory at 8K context.
Discrete GPU: 24 GB of VRAM, of which we assume 92% is usable after driver and context overhead.
Mixture of experts: all 20.915B parameters must be resident in memory even though only ~3.6B are active per token. Memory follows total parameters; speed follows active parameters.
Performance

Estimated performance · recommended MXFP4

Decode, prefill and TTFT below are estimates for MXFP4. We do not have a comparable MXFP4 measurement on this machine.

Estimated decode
~205–295 t/s
Estimated MXFP4; calculated range
Estimated prefill
~1700–3530 t/s
Estimated MXFP4; calculated range
Estimated TTFT at 8K
~2.4–4.9 s
Estimated MXFP4; calculated range
Hardware load reference
600 W
Official spec hardware TDP; not this model run
How the estimate is calculated
  1. Decode: reading 3.6B active parameters at 4.25 bits/weight takes 2.72 ms at 936 GB/s x 75% achieved efficiency.
  2. MoE routing penalty of 15% applied: expert gathers are less bandwidth-efficient than a dense sweep.
  3. Prefill: 142 TFLOPS (FP16) x 32% assumed model-FLOPs utilisation, divided by 2 x 8.7B parameters per token.
  4. Prefill uses 8.7B effective parameters, not the 3.6B 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.

Submit a benchmarkSubmissions stay pending until reviewed.
This model at other quantizations on this machine
QuantizationResidentDiskTotal neededUtilisationFitMax context
MXFP4recommended11.3 GB11.3 GB13.0 GB59%Comfortable128K
BF1639.0 GB39.0 GB41.5 GB188%Does not fit0
Cost of running gpt-oss-20b on this machineUnited States (federal) · C corporation · 8h/day
Monthly economic cost
$26
Calculated after tax
Codex/Claude Code
$100/mo
≈ $100/mo · local is $74 less
Net cash at purchase
$1,981
Calculated VAT not reclaimable
Total over 5 years
$1,578
Calculated after tax, after resale
Cost per USD/1M tokens
$0.83
Calculated 31.6M 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: $1,981 in United States (federal), tax/VAT excluded · estimated · Used market exact-match reference (converted catalogue estimate). This is the localized purchase input; the after-tax economic cost is calculated separately below.

Monthly breakdown

Depreciation
1,545 over 5 years, straight-line to a 436 residual
$25.75
Electricity
31.3 kWh/month at 0.140/kWh
$4.38
Cost of capital
4.0%/yr on 1,208 average capital employed
$4.03
Monthly cost before tax$34.16
Electricity tax shield
Running costs are deductible business expenses
−$0.92
First-year expensing
§179 (100.0%)
−$6.93
Monthly economic cost after tax$26.30

Three different numbers, deliberately

Cash cost
$1,981
Money that leaves the bank account on day one, net of reclaimable VAT.
Accounting depreciation
$25.75/month
$1,545 written down over 5 years to a $436 residual.
After-tax economic cost
$26.30/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
$436 (22%) Assumption
Extrapolated.
Electricity
$0.140/kWh Assumption
Average power draw
178 W Calculated
Load 600 W for 20% of powered hours, idle 72 W for the rest — a machine that is on is not generating tokens the whole time.
Investment allowances
§179 100.0% Official spec
Worth $416 in total — first-year expensing that replaces later tax depreciation.
Cost of capital
$4.03/month Assumption
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
126.0 t/s
Calculated purchase price only
Decode per $100/month
949.2 t/s
Calculated after-tax ownership cost
Tokens per joule
1.41
Calculated same as tokens/s per watt
USD per 1M output tokens
$0.83
Calculated at 20% utilisation
Local versus hosted APIs
Hosted modelUSD/1M outputBreak-evenAPI at your volumeVerdictComparison type
Qwen3.8 Flash
Alibaba Cloud
$0.4217.1M/mo$49Local cheaperDifferent model
DeepSeek-V4.1 Flash
DeepSeek
$0.6014.6M/mo$57Local cheaperDifferent model
DeepSeek-V4 Pro
DeepSeek
$1.983.6M/mo$230Local cheaperDifferent model
DeepSeek V4 Pro (Together)
Together AI
$4.401.2M/mo$671Local cheaperDifferent model
GLM-5.3
Z.ai
$4.401.7M/mo$494Local cheaperDifferent model
Claude Haiku 4.5
Anthropic
$5.002.0M/mo$411Local cheaperDifferent 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 $26/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.