gpt-oss-120b on NVIDIA DGX Spark 128 GB
116.829B (5.1B active) on 128 GB at 273 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
MXFP4
Calculated gguf
Memory required
66.5 GB
Calculated of 105.0 GB usable — 63%
Max practical context
128K
Calculated model supports 128K
Memory budget at 8K context
Model weights ⓘ
63.0 GB Calculated
KV cache ⓘ
0.6 GB Calculated
Runtime overhead ⓘ
2.9 GB Estimated
Total required
66.5 GB Calculated
Usable memory ⓘ
105.0 GB Assumption
Headroom
38.5 GB Calculated
Highest-precision quantization that leaves headroom: uses 63% of usable memory at 8K context.
Unified LPDDR: the accelerator allocates from the same 128 GB pool as the OS. We assume 82% is available to the inference engine.
Mixture of experts: all 116.829B parameters must be resident in memory even though only ~5.1B are active per token. Memory follows total parameters; speed follows active parameters.
PerformanceHigh9/10
Decode (generation)
56.4 t/s
Aggregated range 51.5–60.6 across 4 runs
Prefill (prompt)
1680 t/s
Aggregated range 1,512–1,956
TTFT at 8K
~3.9–8.2 s
Estimated time to first token
Power while generating
170 W
Measured 1.04 tokens/s per watt
Backed by 4 published measurements. Every run is listed below with its source, engine version and context depth.
Evidence
| Type | Engine | Context | Decode | Prefill | Date | Source |
|---|---|---|---|---|---|---|
| Measured (community) | llama.cpp b7db35a7 | 0 | 60.57 t/s | 1,956 t/s | 20 Oct 2025 | github |
| Measured (independent) | llama.cpp | 0 | 58.70 t/s | 1,723 t/s | 25 Oct 2025 | blog |
| Measured (community) | llama.cpp b7db35a7 | 4K | 54.14 t/s | 1,637 t/s | 20 Oct 2025 | github |
| Measured (community) | llama.cpp b7db35a7 | 8K | 51.54 t/s | 1,512 t/s | 20 Oct 2025 | github |
| Measured (community) | llama.cpp b7db35a7 | 16K | 47.45 t/s | 1,307 t/s | 20 Oct 2025 | github |
| Measured (community) | llama.cpp b7db35a7 | 32K | 40.55 t/s | 1,027 t/s | 20 Oct 2025 | github |
- llama-bench at zero context depth: the best case, and the number most often quoted in marketing. Reported as 60.57 +/- 0.25.
- Same run at 4K context depth. Reported as 54.14 +/- 0.08.
- 8K context depth. Reported as 51.54 +/- 0.14.
- 16K context depth. Reported as 47.45 +/- 0.08.
- 32K context depth — decode has fallen 33% from the zero-context figure. This is why the site buckets by context rather than quoting one number.
- Independent reproduction. Slightly below the llama.cpp thread figure, which is normal for a different kernel build.
How the estimate is calculated
- Decode: reading 5.1B active parameters at 4.25 bits/weight takes 14.81 ms at 273 GB/s x 67% achieved efficiency.
- MoE routing penalty of 15% applied: expert gathers are less bandwidth-efficient than a dense sweep.
- Prefill: 125 TFLOPS (FP16) x 4 for native FP4 tensor cores x 0.85 calibrated against measured prefill on this platform x 18% assumed model-FLOPs utilisation, divided by 2 x 24.4B parameters per token.
- Prefill uses 24.4B effective parameters, not the 5.1B 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.
Why this is rated "High" confidence
- Independent third-party benchmark: +3
- 4 measurements: +2
- Sources agree closely (±7%): +2
- Complete run metadata (engine, version, context, source, date): +2
Score 9/10. Aggregated with the median, not the mean, so one outlier cannot move the headline figure.
This model at other quantizations on this machine
| Quantization | Weights | Total needed | Utilisation | Fit | Max context |
|---|---|---|---|---|---|
MXFP4recommended | 63.0 GB | 66.5 GB | 63% | Comfortable | 128K |
BF16 | 217.6 GB | 225.7 GB | 215% | Does not fit | 0 |
Cost of running gpt-oss-120b on this machineUnited States (federal) · C corporation · 8h/day
Monthly economic cost
$52
Calculated after tax
Codex/Claude Code
$100/mo
≈ $100/mo · local is $48 less
Net cash at purchase
$3,873
Calculated VAT not reclaimable
Total over 5 years
$3,091
Calculated after tax, after resale
Cost per USD/1M tokens
$7.21
Calculated 7.1M 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 3,408 over 5 years, straight-line to a 465 residual | $56.80 |
| Electricity 9.5 kWh/month at 0.140/kWh | $1.33 |
| Cost of capital 4.0%/yr on 2,169 average capital employed | $7.23 |
| Monthly cost before tax | $65.36 |
| Electricity tax shield Running costs are deductible business expenses | −$0.28 |
| First-year expensing §179 (100.0%) | −$13.55 |
| Monthly economic cost after tax | $51.52 |
Three different numbers, deliberately
Cash cost
$3,873
Money that leaves the bank account on day one, net of reclaimable VAT.
Accounting depreciation
$56.80/month
$3,408 written down over 5 years to a $465 residual.
After-tax economic cost
$51.52/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
$465 (12%) Assumption
Extrapolated.
Electricity
$0.140/kWh Assumption
Average power draw
54 W Calculated
Load 170 W for 20% of powered hours, idle 25 W for the rest — a machine that is on is not generating tokens the whole time.
Investment allowances
§179 100.0% Official spec
Worth $813 in total — first-year expensing that replaces later tax depreciation.
Cost of capital
$7.23/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
14.6 t/s
Calculated purchase price only
Decode per $100/month
109.5 t/s
Calculated after-tax ownership cost
Tokens per joule
1.04
Calculated same as tokens/s per watt
USD per 1M output tokens
$7.21
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 | 33.5M/mo | $11 | API cheaper | Different model |
| DeepSeek-V4 Flash DeepSeek | $0.66 | 21.3M/mo | $17 | API cheaper | Different model |
| DeepSeek-V4 Pro DeepSeek | $1.98 | 7.1M/mo | $52 | About equal | Different model |
| DeepSeek V4 Pro (Together) Together AI | $4.40 | 2.4M/mo | $152 | Local cheaper | Different model |
| GLM-5.3 Z.ai | $4.40 | 3.3M/mo | $112 | Local cheaper | Different model |
| Claude Haiku 4.5 Anthropic | $5.00 | 4.0M/mo | $93 | 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 $52/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.
gpt-oss-120b on other hardware
| Hardware | Decode | Price | Fit |
|---|---|---|---|
| Quad RTX 5090 workstation (4x 32 GB) | ~339–487 t/s | €13.999 | Comfortable |
| RTX PRO 6000 Blackwell workstation (96 GB) | ~257–369 t/s | €11.499 | Comfortable |
| RTX PRO 6000 Max-Q workstation (96 GB, 300 W) | ~257–369 t/s | €11.999 | Comfortable |
| Mac Studio M5 Ultra 512 GB | ~203–293 t/s | €12.599 | Comfortable |
| Mac Studio M5 Ultra 256 GB | ~203–293 t/s | €9.599 | Comfortable |
| Mac Studio M5 Ultra 96 GB | ~203–293 t/s | €6.599 | Fits |
Nearest alternatives to the NVIDIA DGX Spark 128 GB
gpt-oss-120b on Framework Desktop (Ryzen AI Max+ 395, 128 GB)128 GBgpt-oss-120b on GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB)128 GBgpt-oss-120b on Quad RTX 5090 workstation (4x 32 GB)128 GBgpt-oss-120b on Mac Studio M5 Max 128 GB128 GBgpt-oss-120b on Dual used RTX 4090 workstation (48 GB)48 GB
Compare NVIDIA DGX Spark 128 GB against Framework Desktop (Ryzen AI Max+ 395, 128 GB) →
Compare NVIDIA DGX Spark 128 GB against Framework Desktop (Ryzen AI Max+ 395, 128 GB) →
Page quality score (why this page is or is not indexed)
12/13 — 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 at least one real measurement (+4)
- Has a price (+2)
- Has an ownership economics calculation (+2)
- Well connected (16 internal links) (+1)