Cheapest box to run Gemma-4-31B at BF16
Google’s Gemma-4-31B is 30.7B parameters, dense. The number below is computed from that and each vendor’s published specs, not asserted. Switch the quant or the context and it recomputes in place.
Gemma-4-31B at BF16 needs 64GB.
Cheapest Config that fits: Ascent GX10, GB10 Grace Blackwell, 128GB. Fits, 56GB headroom. Est. 2.9 tok/s.
Best offer today: $3,999.99 at Best Buy — in stock, checked 33 days ago.
Price history: 34 days recorded.
Dense. Only 10 of 60 layers are global attention; the sliding-window layers stop growing past 1024 tokens, so the K/V figure counts the global layers alone.
Offers — Ascent GX10, GB10 Grace Blackwell, 128GB, 1TB
- Best BuyRead on 28 Aug 2026 and not re-read since: Best Buy blocks our automated checks, so this number is old and we say so rather than imply it is current. It was $1,999 under ASUS's own store once ASUS raised its price in September.in stockchecked 33 days ago$3,999.99
- ASUSlist pricein stockchecked 2 hours ago$5,999
- Mac mini, M6, 16GB$1,099Won't fit−51GB headroom
- Mac mini, M6, 24GB$1,299Won't fit−43GB headroom
- Mac mini, M6, 32GB$1,499Won't fit−35GB headroom
- Mac mini, M5 Pro, 24GB$1,699Won't fit−43GB headroom
- Mac mini, M5 Pro, 48GB$2,499Won't fit−20GB headroom
- Mac Studio, M5 Max, 36GB$2,499Won't fit−31GB headroom
- Mac mini, M5 Pro, 64GB$2,899Won't fit−5GB headroom
- GeForce RTX 5090, 32GB$2,999.99+hostWon't fit−33GB headroom
- Won't fit−5GB headroom
- Mac Studio, M5 Max, 48GB$3,099Won't fit−20GB headroom
- Mac Studio, M5 Max, 64GB$3,499Won't fit−5GB headroom
- Fits56GB headroomest. 2.9 tok/s
- Fits24GB headroomest. 8.7 tok/s
- Fits56GB headroomest. 5.8 tok/s
- Fits56GB headroomest. 2.9 tok/s
- Fits56GB headroomest. 6.5 tok/s
- Fits24GB headroomest. 12.7 tok/s
- Fits184GB headroomest. 8.7 tok/s
- Fits184GB headroomest. 12.7 tok/s
- Fits440GB headroomest. 12.7 tok/s
| Config | Memory | Price | $/GB | Fit | Headroom | Est. speed |
|---|---|---|---|---|---|---|
| Mac mini, M6 | 16GB | $1,099 | $69 | Won't fit | −51GB | — |
| Mac mini, M6 | 24GB | $1,299 | $54 | Won't fit | −43GB | — |
| Mac mini, M6 | 32GB | $1,499 | $47 | Won't fit | −35GB | — |
| Mac mini, M5 Pro | 24GB | $1,699 | $71 | Won't fit | −43GB | — |
| Mac mini, M5 Pro | 48GB | $2,499 | $52 | Won't fit | −20GB | — |
| Mac Studio, M5 Max | 36GB | $2,499 | $69 | Won't fit | −31GB | — |
| Mac mini, M5 Pro | 64GB | $2,899 | $45 | Won't fit | −5GB | — |
| GeForce RTX 5090 | 32GB | $2,999.99+host | $94 | Won't fit | −33GB | — |
| Mac Studio, M4 Max, refurb | 64GB | $3,049 | $48 | Won't fit | −5GB | — |
| Mac Studio, M5 Max | 48GB | $3,099 | $65 | Won't fit | −20GB | — |
| Mac Studio, M5 Max | 64GB | $3,499 | $55 | Won't fit | −5GB | — |
| Ascent GX10, GB10 Grace Blackwell | 128GB | $3,999.99 | $31 | Fits | 56GB | 2.9 tok/s |
| Mac Studio, M3 Ultra, refurb | 96GB | $4,329 | $45 | Fits | 24GB | 8.7 tok/s |
| Mac Studio, M4 Max, refurb | 128GB | $4,409 | $34 | Fits | 56GB | 5.8 tok/s |
| DGX Spark, GB10 Grace Blackwell | 128GB | $4,699 | $37 | Fits | 56GB | 2.9 tok/s |
| Mac Studio, M5 Max | 128GB | $5,099 | $40 | Fits | 56GB | 6.5 tok/s |
| Mac Studio, M5 Ultra | 96GB | $5,499 | $57 | Fits | 24GB | 12.7 tok/s |
| Mac Studio, M3 Ultra, refurb | 256GB | $7,729 | $30 | Fits | 184GB | 8.7 tok/s |
| Mac Studio, M5 Ultra | 256GB | $9,499 | $37 | Fits | 184GB | 12.7 tok/s |
| Mac Studio, M5 Ultra | 512GB | — | — | Fits | 440GB | 12.7 tok/s |
GeForce RTX 5090 is a card, not a computer: every price shown for it is the cheapest listing we found plus $1,000 for the minimum host machine.
memory = 30.7B × 2.000 B/param (16 bits) + 0.168 GB/1K × 8K K/V + 1.2 GB runtime
speed = bandwidth × 0.65 ÷ (30.7B active × 2.000 B/param) = 61.4 GB read per token
usable = memory − operating system reserve (8% on a Mac, 6% on a DGX Spark, 4% on a card)
Bandwidth is the vendor’s published figure. Every speed is an estimate and says so. Methodology.
What else Ascent GX10 GB10 Grace Blackwell 128GB runs
- Qwen3-235B-A22BQ3_K_M, est. 12.7 tok/s
- MiniMax-M2Q3_K_M, est. 27.9 tok/s
- Mistral-Large-2411Q6_K, est. 1.8 tok/s
- gpt-oss-120bQ6_K, est. 32.6 tok/s
- Llama-4-Scout-17B-16EQ8_0, est. 7.6 tok/s
- Qwen3-Coder-NextQ8_0, est. 42.8 tok/s
Best quant that still fits at 8K context, largest model first. Full list on the Config page.
Bandwidth 273GB/s, usable memory 120GB of 128GB.