Cheapest box to run Llama-3.3-70B at BF16
Meta’s Llama-3.3-70B is 70B 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.
Llama-3.3-70B at BF16 needs 144GB.
Cheapest Config that fits: Mac Studio, M5 Ultra, 256GB. Fits, 104GB headroom. Est. 5.6 tok/s.
Best offer today: $9,499 at Apple — in stock, checked 10 hours ago.
Collecting price history — 2 of 90 days. No timing verdict yet.
Offers — Mac Studio, M5 Ultra, 256GB, 1TB
- Mac mini, M6, 16GB$1,099Won't fit−131GB headroom
- Mac mini, M6, 24GB$1,299Won't fit−123GB headroom
- Mac mini, M6, 32GB$1,499Won't fit−115GB headroom
- Mac mini, M5 Pro, 24GB$1,699Won't fit−123GB headroom
- Mac mini, M5 Pro, 48GB$2,499Won't fit−100GB headroom
- Mac Studio, M5 Max, 36GB$2,499Won't fit−111GB headroom
- Mac mini, M5 Pro, 64GB$2,899Won't fit−85GB headroom
- GeForce RTX 5090, 32GB$2,999.99+hostWon't fit−113GB headroom
- Mac Studio, M5 Max, 48GB$3,099Won't fit−100GB headroom
- Mac Studio, M5 Max, 64GB$3,499Won't fit−85GB headroom
- Won't fit−24GB headroom
- Won't fit−24GB headroom
- Won't fit−24GB headroom
- Won't fit−56GB headroom
- Fits104GB headroomest. 5.6 tok/s
- Fits360GB headroomest. 5.6 tok/s
| Config | Memory | Price | $/GB | Fit | Headroom | Est. speed |
|---|---|---|---|---|---|---|
| Mac mini, M6 | 16GB | $1,099 | $69 | Won't fit | −131GB | — |
| Mac mini, M6 | 24GB | $1,299 | $54 | Won't fit | −123GB | — |
| Mac mini, M6 | 32GB | $1,499 | $47 | Won't fit | −115GB | — |
| Mac mini, M5 Pro | 24GB | $1,699 | $71 | Won't fit | −123GB | — |
| Mac mini, M5 Pro | 48GB | $2,499 | $52 | Won't fit | −100GB | — |
| Mac Studio, M5 Max | 36GB | $2,499 | $69 | Won't fit | −111GB | — |
| Mac mini, M5 Pro | 64GB | $2,899 | $45 | Won't fit | −85GB | — |
| GeForce RTX 5090 | 32GB | $2,999.99+host | $94 | Won't fit | −113GB | — |
| Mac Studio, M5 Max | 48GB | $3,099 | $65 | Won't fit | −100GB | — |
| Mac Studio, M5 Max | 64GB | $3,499 | $55 | Won't fit | −85GB | — |
| Ascent GX10, GB10 Grace Blackwell | 128GB | $3,999.99 | $31 | Won't fit | −24GB | — |
| DGX Spark, GB10 Grace Blackwell | 128GB | $4,699 | $37 | Won't fit | −24GB | — |
| Mac Studio, M5 Max | 128GB | $5,099 | $40 | Won't fit | −24GB | — |
| Mac Studio, M5 Ultra | 96GB | $5,499 | $57 | Won't fit | −56GB | — |
| Mac Studio, M5 Ultra | 256GB | $9,499 | $37 | Fits | 104GB | 5.6 tok/s |
| Mac Studio, M5 Ultra | 512GB | — | — | Fits | 360GB | 5.6 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 = 70B × 2.000 B/param (16 bits) + 0.328 GB/1K × 8K K/V + 1.2 GB runtime
speed = bandwidth × 0.65 ÷ (70B active × 2.000 B/param) = 140.0 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 Mac Studio M5 Ultra 256GB runs
- Llama-4-Maverick-17B-128EQ4_K_M, est. 58.5 tok/s
- Qwen3-235B-A22BQ6_K, est. 33.3 tok/s
- Mistral-Large-2411Q8_0, est. 6.0 tok/s
- gpt-oss-120bBF16, est. 58.8 tok/s
- Llama-4-Scout-17B-16EBF16, est. 17.6 tok/s
- Llama-3.3-70BBF16, est. 5.6 tok/s
Best quant that still fits at 8K context, largest model first. Full list on the Config page.
Bandwidth 1.2TB/s, usable memory 248GB of 256GB.