Framework Desktop (Ryzen AI Max+ 395, 128 GB)

An APU with a 256-bit LPDDR5X bus and 40 RDNA 3.5 compute units, sold in mini-PCs with up to 128 GB of shared memory. The cheapest way to put a 100B-class MoE model entirely in memory, at roughly a third of the bandwidth of a Mac Studio Ultra.

AMDcomponent17 models fit1 measured
Memory
128 GB
Official spec LPDDR5X-8000 unified (soldered)
Memory bandwidth
256 GB/s
Official spec ~48% achieved in practice
Specifications
CPU
AMD Ryzen AI Max+ 395, 16-core Zen 5 Official spec
GPU
Radeon 8060S, 40 CU RDNA 3.5 Official spec
Memory
128 GB LPDDR5X-8000 unified (soldered) Official spec
Memory bandwidth
256 GB/s Official spec source
Achieved bandwidth
~48% (123 GB/s) Measured
FP16 compute
~59 TFLOPS Official spec
Architecture
Unified LPDDR (integrated accelerator) Official spec
Nominal power
140 W Official spec
Measured load power
130 W Measured
Idle power
12 W Measured
Released
1 Aug 2025 Official spec
128 GB of unified memory for under EUR 2,500 — by a wide margin the cheapest way to hold a 100B-class MoE model in memory. Achieved bandwidth on ROCm is materially below the 256 GB/s theoretical figure, which is why community results for the same model vary so much. Achieved-bandwidth coefficient CALIBRATED to 48% from 6 measured runs on this configuration.
Model performance on Framework Desktop (Ryzen AI Max+ 395, 128 GB)1 measured, 16 estimated
17 of 17 rows
ModelQuantMemoryDecodePrefillContextFitConfidence
gpt-oss-20b
OpenAI · 20.915B (3.6B active)
MXFP413.0 GB~41.6–59.8 t/s~398–826 t/s128KComfortableEstimated
gpt-oss-120b
OpenAI · 116.829B (5.1B active)
MXFP466.5 GB32.5 t/s128KComfortableMedium
Qwen3-Coder 30B-A3B
Alibaba Qwen · 30.532B (3.3B active)
Q8_033.5 GB~23.5–33.8 t/s~344–714 t/s256KComfortableEstimated
Qwen3 30B-A3B
Alibaba Qwen · 30.532B (3.3B active)
Q8_033.5 GB~23.5–33.8 t/s~344–714 t/s40KComfortableEstimated
Gemma 4 26B-A4B
Google DeepMind · 25.806B (3.8B active)
Q8_029.7 GB~20.5–29.5 t/s~349–724 t/s256KComfortableEstimated
Qwen3.5 122B-A10B
Alibaba Qwen · 125.086B (10B active)
Q4_K_M76.7 GB~13.8–19.9 t/s~97.6–203 t/s192KComfortableEstimated
Qwen3 8B
Alibaba Qwen · 8.191B
Q8_010.6 GB~11.4–16.4 t/s~421–875 t/s40KComfortableEstimated
Phi-4 14B
Microsoft · 14.66B
Q8_017.8 GB~6.4–9.2 t/s~235–489 t/s16KComfortableEstimated
Devstral Small 24B
Mistral AI · 23.572B
Q8_026.8 GB~4–5.8 t/s~146–304 t/s128KComfortableEstimated
Mistral Small 3.2 24B
Mistral AI · 24.011B
Q8_027.2 GB~3.9–5.7 t/s~144–299 t/s128KComfortableEstimated
Gemma 3 27B
Google DeepMind · 27.432B
Q8_033.4 GB~3.5–5 t/s~126–261 t/s128KComfortableEstimated
Qwen3.8 27B
Alibaba Qwen · 27.781B
Q8_031.9 GB~3.4–4.9 t/s~124–258 t/s256KComfortableEstimated
Qwen3.6 27B
Alibaba Qwen · 27.781B
Q8_031.9 GB~3.4–4.9 t/s~124–258 t/s256KComfortableEstimated
Gemma 4 31B
Google DeepMind · 31.273B
Q8_041.0 GB~3–4.4 t/s~110–229 t/s64KComfortableEstimated
Qwen3 32B
Alibaba Qwen · 32.762B
Q8_037.1 GB~2.9–4.2 t/s~105–219 t/s40KComfortableEstimated
DeepSeek-R1-Distill 32B
DeepSeek · 32.764B
Q8_037.1 GB~2.9–4.2 t/s~105–219 t/s128KComfortableEstimated
Llama 3.3 70B
Meta · 70.554B
Q8_076.8 GB~1.3–1.9 t/s~48.9–102 t/s64KComfortableEstimated
Rows in grey are estimates from our bandwidth model, not measurements — they are shown as a range and never as a precise figure. Use the “measured only” filter to see just the 1 pairing on this machine that a real benchmark backs.
Ownership economicsUnited States (federal) · C corporation · 8h/day
Monthly economic cost
$28
Calculated after tax
Codex/Claude Code
$100/mo
≈ $100/mo · local is $72 less
Net cash at purchase
$2,161
Calculated VAT not reclaimable
Total over 5 years
$1,673
Calculated after tax, after resale
Cost per USD/1M tokens
$4.34
Calculated 6.4M 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
1,837 over 5 years, straight-line to a 324 residual
$30.62
Electricity
6.3 kWh/month at 0.140/kWh
$0.88
Cost of capital
4.0%/yr on 1,243 average capital employed
$4.14
Monthly cost before tax$35.63
Electricity tax shield
Running costs are deductible business expenses
−$0.18
First-year expensing
§179 (100.0%)
−$7.56
Monthly economic cost after tax$27.89

Three different numbers, deliberately

Cash cost
$2,161
Money that leaves the bank account on day one, net of reclaimable VAT.
Accounting depreciation
$30.62/month
$1,837 written down over 5 years to a $324 residual.
After-tax economic cost
$27.89/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
$324 (15%) Assumption
Two GPU generations later, the card is worth a fraction of its list price.
Electricity
$0.140/kWh Assumption
Average power draw
36 W Calculated
Load 130 W for 20% of powered hours, idle 12 W for the rest — a machine that is on is not generating tokens the whole time.
Investment allowances
§179 100.0% Official spec
Worth $454 in total — first-year expensing that replaces later tax depreciation.
Cost of capital
$4.14/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.