EDGE OSOPEN-WEIGHT LLM COMPARISON
DeepSeek-V3.2 vs DeepSeek-V4-FlashSpecs, VRAM and the hardware to run them
DeepSeek-V4-Flash is the newer, smaller design: 284B parameters with 13B active and a 1M-token context, versus 685B with 37B active and 128K for V3.2. Flash needs less than half the memory.
Parameters
685B vs 284B
Active per token
37B vs 13B
Context
128K vs 1M
Memory, 1 user
419 GB vs 177 GB
Which should you choose?
CHOOSE DEEPSEEK-V3.2 IF
You have already validated V3.2 in production and only need a 128K context.
CHOOSE DEEPSEEK-V4-FLASH IF
You are starting now: it is newer, needs far less memory, has an 8× longer context and generates tokens faster.
Still deciding between DeepSeek-V3.2 and DeepSeek-V4-Flash?
Tell us your models, users and data. We will recommend a setup and run it inside your own building with Vanik Edge OS. Free, no obligation.
DeepSeek-V3.2 vs DeepSeek-V4-Flash specs
From each developer’s model card. Green marks the lighter figure where less is better for running cost.
| Specification | DeepSeek-V3.2 | DeepSeek-V4-Flash |
|---|---|---|
| Developer | DeepSeek | DeepSeek |
| Total parameters | 685B | 284B |
| Active per token | 37B | 13B |
| Design | Mixture of experts | Mixture of experts |
| Experts | 256 routed experts + 1 shared, 8 active per token | Mixture of experts (count not stated) |
| Context window | 128K | 1M tokens |
| Input and output | Text | Text input, text output |
| Layers | 61 layers | Not stated |
| Attention | Multi-head latent attention with DeepSeek Sparse Attention | Hybrid Compressed Sparse Attention and Heavily Compressed Attention |
| License | MIT | MIT |
| Released weight formats | FP8, NVFP4 (4-bit), BF16 | FP4 + FP8 (native) |
| Weights at NVFP4 / FP4 + FP8 | About 397 GB | About 168 GB |
VRAM needed to run them
GPU memory at the lowest released precision. Each user needs their own KV cache, so memory grows with users and context length.
| GPU memory needed (NVFP4 / FP4 + FP8) | DeepSeek-V3.2 | DeepSeek-V4-Flash |
|---|---|---|
| 1 user, 8K context | About 419 GB | About 177 GB |
| 1 user, 128K context | About 428 GB | About 178 GB |
| 16 users, 8K context each | About 428 GB | About 178 GB |
| 64 users, 8K context each | About 459 GB | About 183 GB |
Which GPUs can run DeepSeek-V3.2 and DeepSeek-V4-Flash
Smallest setup on each system, with 8K context per user. Green fits on one GPU or system, amber needs several in one server, grey needs more than one 8-GPU server.
| Hardware | DeepSeek-V3.2: 1 user | DeepSeek-V3.2: 16 users | DeepSeek-V4-Flash: 1 user | DeepSeek-V4-Flash: 16 users |
|---|---|---|---|---|
| NVIDIA DGX Spark 128 GB per system | No | No | 2 systems | 2 systems |
| NVIDIA L40S 48 GB per GPU | No | No | 8 GPUs | 8 GPUs |
| NVIDIA RTX PRO 6000 Blackwell 96 GB per GPU | 8 GPUs | 8 GPUs | 4 GPUs | 4 GPUs |
| NVIDIA H100 NVL 94 GB per GPU | 8 GPUs | 8 GPUs | 4 GPUs | 4 GPUs |
| NVIDIA H200 NVL 141 GB per GPU | 4 GPUs | 4 GPUs | 2 GPUs | 2 GPUs |
| AMD Instinct MI300X 192 GB per GPU | 4 GPUs | 4 GPUs | 2 GPUs | 2 GPUs |
| AMD Instinct MI325X 256 GB per GPU | 2 GPUs | 2 GPUs | Yes, 1 GPU | Yes, 1 GPU |
| NVIDIA DGX B200 1.44 TB per system | Yes, 1 system | Yes, 1 system | Yes, 1 system | Yes, 1 system |
Setups in amber or grey need several GPUs or servers working as one. Buying the GPUs is the easy part. Vanik Edge OS splits the model across them, keeps it running if a server fails and manages it all from one console. See how Edge OS clusters GPU servers
Which is faster?
On the same hardware, DeepSeek-V4-Flash should generate tokens about 2.8× faster for each user: it reads about 7.7 GB of weights per token (13B active at FP4 + FP8) versus 21 GB for DeepSeek-V3.2. Token generation is limited mostly by memory bandwidth, so fewer active bytes means faster replies. Real throughput also depends on batching, the inference engine and the GPU.
FROM SIZING TO RUNNING
Sizing the hardware is the easy part.
Getting several GPUs to work as one system for a whole company is where most projects slow down. That is the part Vanik Edge OS does for you.
STEP 01 · THIS TOOL
Size it
Work out the GPU memory your models need and which GPUs or servers can hold them.
STEP 02 · VANIK EDGE OS
Cluster it
Edge OS joins your GPU servers into one cluster, splits a large model across them and keeps it running if a server fails. NVIDIA and AMD servers work side by side.
STEP 03 · VANIK EDGE OS
Run it
Your team gets AI apps with single sign-on and an audit trail, managed from one console, with your data staying inside your building.
DeepSeek-V3.2 vs DeepSeek-V4-Flash questions
What is the difference between DeepSeek-V3.2 and DeepSeek-V4-Flash?
DeepSeek-V4-Flash is the newer, smaller design: 284B parameters with 13B active and a 1M-token context, versus 685B with 37B active and 128K for V3.2. Flash needs less than half the memory.
DeepSeek-V3.2 vs DeepSeek-V4-Flash: which is better to run on-premise?
Choose DeepSeek-V3.2 if you have already validated V3.2 in production and only need a 128K context. Choose DeepSeek-V4-Flash if you are starting now: it is newer, needs far less memory, has an 8× longer context and generates tokens faster.
Which needs less VRAM, DeepSeek-V3.2 or DeepSeek-V4-Flash?
DeepSeek-V4-Flash. For one user with 8K context it needs about 177 GB, versus about 419 GB for DeepSeek-V3.2. For 16 users it needs about 178 GB versus 428 GB.
Which is faster, DeepSeek-V3.2 or DeepSeek-V4-Flash?
On the same hardware, DeepSeek-V4-Flash should generate tokens about 2.8× faster for each user: it reads about 7.7 GB of weights per token (13B active at FP4 + FP8) versus 21 GB for DeepSeek-V3.2. Token generation is limited mostly by memory bandwidth, so fewer active bytes means faster replies.
Can I run DeepSeek-V3.2 locally?
Yes. At NVFP4 it needs about 419 GB of GPU memory for one user with 8K context, which fits on 2 × AMD Instinct MI325X. A team of 16 needs about 428 GB: 2 × AMD Instinct MI325X.
Can I run DeepSeek-V4-Flash locally?
Yes. At FP4 + FP8 it needs about 177 GB of GPU memory for one user with 8K context, which fits on 2 × NVIDIA DGX Spark. A team of 16 needs about 178 GB: 2 × NVIDIA DGX Spark.
Can I use DeepSeek-V3.2 and DeepSeek-V4-Flash commercially?
DeepSeek-V3.2 is released under MIT. DeepSeek-V4-Flash is released under MIT. Check each license before a commercial deployment.
How are these numbers calculated?
Memory figures come from our VRAM calculator method: weights at the lowest released precision, plus the KV cache for each user’s context, plus about 5 to 10% runtime overhead, keeping 10% of each GPU free. Where a developer does not publish the cache size we estimate it from the architecture. These are planning estimates, not benchmarks.
Related comparisons
Sources: NVIDIA: DeepSeek-V3.2 model card; Hugging Face: deepseek-ai/DeepSeek-V4-Flash-0731; Hugging Face blog: DeepSeek-V4. Memory figures are planning estimates; real use depends on the inference engine and settings. Last checked November 12, 2026.