Cheapest box to run DeepSeek-V4-Flash at Q3_K_M
DeepSeek’s DeepSeek-V4-Flash is 284B parameters, 13B of them read per token. 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.
DeepSeek-V4-Flash at Q3_K_M needs 140GB.
Cheapest Config that fits: Mac Studio, M3 Ultra, 256GB, refurbished. Fits, 108GB headroom. Est. 64.5 tok/s.
Best offer today: $7,729 at Apple Certified Refurbished — in stock, checked 2 hours ago.
Price history: 34 days recorded.
Compressed sparse attention. Each layer stores one latent, divided by the compression ratio in its config (mostly 4x or 128x); the figure matches DeepSeek's own claim of about a tenth of the V3 cache.
Offers — Mac Studio, M3 Ultra, 256GB, 1TB, refurbished
- Apple Certified RefurbishedRefurbished by Apple, with Apple's one-year limited warranty and 14-day returns. Apple says refurbished supply is usually very limited and often runs out, so this row can go dead faster than a new one.in stockchecked 2 hours ago$7,729
- Mac mini, M6, 16GB$1,099Won't fit−127GB headroom
- Mac mini, M6, 24GB$1,299Won't fit−119GB headroom
- Mac mini, M6, 32GB$1,499Won't fit−111GB headroom
- Mac mini, M5 Pro, 24GB$1,699Won't fit−119GB headroom
- Mac mini, M5 Pro, 48GB$2,499Won't fit−96GB headroom
- Mac Studio, M5 Max, 36GB$2,499Won't fit−107GB headroom
- Mac mini, M5 Pro, 64GB$2,899Won't fit−81GB headroom
- GeForce RTX 5090, 32GB$2,999.99+hostWon't fit−109GB headroom
- Won't fit−81GB headroom
- Mac Studio, M5 Max, 48GB$3,099Won't fit−96GB headroom
- Mac Studio, M5 Max, 64GB$3,499Won't fit−81GB headroom
- Won't fit−20GB headroom
- Won't fit−52GB headroom
- Won't fit−20GB headroom
- Won't fit−20GB headroom
- Won't fit−20GB headroom
- Won't fit−52GB headroom
- Fits108GB headroomest. 64.5 tok/s
- Fits108GB headroomest. 94.4 tok/s
- Fits364GB headroomest. 94.4 tok/s
| Config | Memory | Price | $/GB | Fit | Headroom | Est. speed |
|---|---|---|---|---|---|---|
| Mac mini, M6 | 16GB | $1,099 | $69 | Won't fit | −127GB | — |
| Mac mini, M6 | 24GB | $1,299 | $54 | Won't fit | −119GB | — |
| Mac mini, M6 | 32GB | $1,499 | $47 | Won't fit | −111GB | — |
| Mac mini, M5 Pro | 24GB | $1,699 | $71 | Won't fit | −119GB | — |
| Mac mini, M5 Pro | 48GB | $2,499 | $52 | Won't fit | −96GB | — |
| Mac Studio, M5 Max | 36GB | $2,499 | $69 | Won't fit | −107GB | — |
| Mac mini, M5 Pro | 64GB | $2,899 | $45 | Won't fit | −81GB | — |
| GeForce RTX 5090 | 32GB | $2,999.99+host | $94 | Won't fit | −109GB | — |
| Mac Studio, M4 Max, refurb | 64GB | $3,049 | $48 | Won't fit | −81GB | — |
| Mac Studio, M5 Max | 48GB | $3,099 | $65 | Won't fit | −96GB | — |
| Mac Studio, M5 Max | 64GB | $3,499 | $55 | Won't fit | −81GB | — |
| Ascent GX10, GB10 Grace Blackwell | 128GB | $3,999.99 | $31 | Won't fit | −20GB | — |
| Mac Studio, M3 Ultra, refurb | 96GB | $4,329 | $45 | Won't fit | −52GB | — |
| Mac Studio, M4 Max, refurb | 128GB | $4,409 | $34 | Won't fit | −20GB | — |
| DGX Spark, GB10 Grace Blackwell | 128GB | $4,699 | $37 | Won't fit | −20GB | — |
| Mac Studio, M5 Max | 128GB | $5,099 | $40 | Won't fit | −20GB | — |
| Mac Studio, M5 Ultra | 96GB | $5,499 | $57 | Won't fit | −52GB | — |
| Mac Studio, M3 Ultra, refurb | 256GB | $7,729 | $30 | Fits | 108GB | 64.5 tok/s |
| Mac Studio, M5 Ultra | 256GB | $9,499 | $37 | Fits | 108GB | 94.4 tok/s |
| Mac Studio, M5 Ultra | 512GB | — | — | Fits | 364GB | 94.4 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 = 284B × 0.489 B/param (3.91 bits) + 0.006 GB/1K × 8K K/V + 1.2 GB runtime
speed = bandwidth × 0.5 ÷ (13B active × 0.489 B/param) = 6.4 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 Mac Studio M3 Ultra 256GB refurb runs
- Qwen3-Coder-480B-A35B-InstructQ3_K_M, est. 23.9 tok/s
- Llama-4-Maverick-17B-128EQ4_K_M, est. 39.9 tok/s
- GLM-4.6Q4_K_M, est. 21.2 tok/s
- DeepSeek-V4-FlashQ6_K, est. 38.4 tok/s
- Qwen3-235B-A22BQ6_K, est. 22.7 tok/s
- MiniMax-M2Q8_0, est. 38.5 tok/s
Best quant that still fits at 8K context, largest model first. Full list on the Config page.
Bandwidth 819GB/s, usable memory 248GB of 256GB.