DeepSeek-V4 Flash
DeepSeek describes the core MoE as 284B total with 13B active; the pinned weight archive contains about 304B tensors including auxiliary weights. Its single 512-dimensional KV head keeps long-context cache growth comparatively small.
DeepSeekMixture of expertscodingreasoning8 machines can run it
Total parameters
304.18B
Official spec sets memory need
Active parameters
13B
Official spec sets decode speed
Context
1M
Official spec tokens
4-bit weights
81 GB
Calculated before KV cache
This is a sparse mixture of experts. All 304.18B of weights must be available to the runtime, but only 13B are read per generated token. Ordinary runtimes keep the full quantized model in memory; a runtime-specific SSD-streaming artifact can retain a smaller working set and fetch expert data on demand, trading speed for capacity.
Model specification
Publisher
DeepSeek Official spec
Architecture
Sparse mixture of experts Official spec
Total parameters
304.18B Official spec
Active parameters
13B per token Official spec
Context length
1,048,576 tokens Official spec
Attention ⓘ
MQA · 1 KV heads × 512 dim × 43 layers Official spec
Licence
MIT Official spec source
Specification confidence ⓘ
High · verified 2026-10-19 Official spec source
Released
1 May 2026 Official spec
Official source
Quantizations and memory
| Quantization | Format | Bits/weight | Resident | Download | Quality kept |
|---|---|---|---|---|---|
DS4 Q2 (mixed) | gguf | 2.29 | 81.0 GB | 81.0 GB | — |
MLX 4-bit | mlx | 4.5 | 165.7 GB | 165.7 GB | 98.0% |
Q4_K_Mdefault | gguf | 4.85 | 176.9 GB | 176.9 GB | 98.5% |
Q5_K_M | gguf | 5.69 | 207.5 GB | 207.5 GB | 99.3% |
Q8_0 | gguf | 8.5 | 307.0 GB | 307.0 GB | 99.9% |
Generic weight sizes are computed from the parameter count and bits per weight plus a format-specific overhead. Runtime-specific artifacts use their published resident and download footprints; streamed models can therefore require much more disk than memory. Quality retention is an assumption, not a measured evaluation.
Coding & quality benchmarksCompare coding results →
No published benchmark results have been imported for this model yet. This is missing evidence, not a score of zero.
We show source metrics rather than deriving one opaque quality number. Different benchmarks measure genuinely different things, and collapsing them into a single score would hide exactly the disagreements worth seeing.
Recommendations
Cheapest that can run it
18.1 t/s · $3,873
Lowest purchase price among configurations where the model fits at some quantization in our catalogue. Speed is not considered.
Cheapest above 20 t/s
~51.8–74.5 t/s · $5,674
20 tokens/second is roughly the point at which generation keeps pace with reading. Below it, interactive use feels like waiting.
Cheapest above 40 t/s
~51.8–74.5 t/s · $5,674
40 tokens/second is the threshold most people describe as comfortable for coding agents, where output arrives faster than you can review it.
Fastest with real measurements
18.05 t/s measured at DS4 Q2 (mixed) · $3,873
Highest throughput among configurations with an actual published measurement rather than our estimate.
Best throughput per purchase-price unit
~89.3–129 t/s · $8,647
Highest decode tokens/second per 1,000 units of the displayed purchase currency. Ignores running costs and resale — see the economics section for the full picture.
Hardware that runs DeepSeek-V4 Flash1 measured, 7 estimated
8 of 8 rows
| Hardware↕ | Quant | Memory↕ | Decode▼ | Prefill↕ | Context | Price↕ | Fit | Confidence↕ |
|---|---|---|---|---|---|---|---|---|
| NVIDIA DGX H200 (8x H200, 1,128 GB) NVIDIA · 1,128 GB · 38,400 GB/s | Q8_0recommended | 320.7 GB | ~411–592 t/s | ~42550–88370 t/s | 1024K | $359,430 | Comfortable | Estimated |
| Lenovo ThinkStation PX (4x RTX PRO 6000, 384 GB) Lenovo · 384 GB · 7,168 GB/s | Q5_K_Mrecommended | 216.6 GB | ~150–216 t/s | ~2040–4240 t/s | 1024K | $58,553 | Comfortable | Estimated |
| Mac Studio M5 Ultra 256 GB Apple · 256 GB · 1,200 GB/s | MLX 4-bitrecommended | 172.4 GB | ~89.3–129 t/s | ~462–959 t/s | 256K | $8,647 | Comfortable | Estimated |
| Mac Studio M3 Ultra 256 GB Apple · 256 GB · 819 GB/s | MLX 4-bitrecommended | 172.4 GB | ~51.8–74.5 t/s | ~111–231 t/s | 256K | $5,674 | Comfortable | Estimated |
| NVIDIA DGX Station GB300 (748 GB) NVIDIA · 748 GB · 7,100 GB/s | Q8_0recommended | 317.9 GB | ~50.4–72.6 t/s | ~5690–11810 t/s | 1024K | $90,082 | Comfortable | Estimated |
| Mac Studio M5 Ultra 512 GB Apple · 512 GB · 1,200 GB/s | Q8_0recommended | 317.9 GB | ~49.8–71.7 t/s | ~462–959 t/s | 1024K | $11,350 | Comfortable | Estimated |
| Mac Studio M3 Ultra 512 GB Apple · 512 GB · 819 GB/s | Q8_0recommended | 317.9 GB | ~28.3–40.7 t/s | ~111–231 t/s | 1024K | $8,287 | Comfortable | Estimated |
| NVIDIA DGX Spark 128 GB NVIDIA · 128 GB · 273 GB/s | DS4 Q2 (mixed)recommended | 85.1 GB | 18.1 t/s | 826 t/s | 192K | $3,873 | Fits | Medium |
Submit a benchmarkContributions are reviewed before publication.
Similar models
DeepSeek Coder V2 236B-A21B
DeepSeek · 235.741B (21B active) · DeepSeek
GLM-5.3 Flash
Z.ai · 321.323B (18B active) · MIT
DeepSeek Coder V2 Lite 16B-A2.4B
DeepSeek · 15.706B (2.4B active) · DeepSeek
GLM-5.3
Z.ai · 753.33B (40B active) · MIT
DeepSeek V2 236B-A21B
DeepSeek · 235.741B (21B active) · DeepSeek
DeepSeek V2.5 236B-A21B
DeepSeek · 235.741B (21B active) · DeepSeek