DeepSeek V4.1 Flash 552B vs DeepSeek V4 Flash 0731 284B
Side-by-side VRAM requirements, benchmark scores, and GPU compatibility for local AI inference.
Quick verdict
DeepSeek V4 Flash 0731 284B is more hardware-efficient: it needs 108.5 GB at its UD-Q2_K_XL build vs 571.6 GB for DeepSeek V4.1 Flash 552B's Q4_K_M, fitting on 16 GPUs natively.
Analysis
DeepSeek V4.1 Flash replaces V4 Flash 0731 as DeepSeek's small, cheap 'Flash' checkpoint, released six weeks apart. This isn't a re-post-train the way 0731 was over the original April preview: DeepSeek calls the new design a Causal Encoder-Decoder, a real architecture change, and it shows up directly in benchmarks and KV-cache size. The catch is availability: V4 Flash 0731 has a real, vetted quant ladder people can download today; V4.1 Flash, released the same day this comparison was written, does not.
V4.1 Flash is the bigger model on paper, 552B total parameters against 284B, and reads more per token on decode, 16B active against 13B, but the KV-cache story runs the other way: DeepSeek's own figures put its cache at roughly a quarter of V4 Flash 0731's, and this site's math confirms it at the full shared 1,048,576-token context, 2.3 GB against 9.2 GB, on top of a cache that was already unusually small for a frontier-scale MoE. Benchmarks split by category: GPQA Diamond is essentially flat (90.9 against 91), but Terminal-Bench 2.1 jumped from 82.7 to 90.6, DeepSeek's own reported figures for each. Both ship under MIT with no usage restrictions. Modality is a real gain for the new release: V4.1 Flash accepts image input natively, where V4 Flash 0731 is text only. Weights are where the comparison turns sharply in the older model's favor. V4 Flash 0731's routed experts ship natively in a low-bit format, so Unsloth's real GGUF builds run 82.5 GB to 161.9 GB, with the recommended UD-Q2_K_XL at 96.8 GB. V4.1 Flash has no build below its native ~510 GB release yet, so this site's quantization table floors every rung at that same number rather than guessing at a smaller file nobody has published.
Bottom line: If you actually want to run one of these locally today, V4 Flash 0731 is the only real option: a 128 GB-class unified-memory machine holds it comfortably, at the full million-token context. V4.1 Flash is the better model by nearly every other measure, a real architectural advance, a noticeably smaller KV cache, a large Terminal-Bench jump, and native vision, but none of that is reachable outside DeepSeek's own API until a community requantization exists. For agentic coding work specifically and a willingness to use the API either way, V4.1 Flash is the one to point traffic at; for actually holding the weights on your own hardware, it's V4 Flash 0731, at least for now.
Total vs. active parameters: a bigger model that reads more per token
Both models are MoE, so total parameters (the VRAM bill) and active parameters (the compute/speed bill) move independently. V4.1 Flash is larger on both axes.
The KV cache shrank again, six weeks later
V4 Flash 0731 already compressed its KV cache far below a conventional attention stack. V4.1 Flash's new Causal Encoder-Decoder design compresses it again, on top of that, at the same full 1,048,576-token context both models support.
KV cache only, at FP16. Weights and activation overhead sit on top of these figures.
At the full 1,048,576-token window, V4.1 Flash's KV cache is 2.3 GB, against V4 Flash 0731's already-small 9.2 GB, a roughly 4x reduction on top of a design that was already far ahead of a conventional dense-attention model at this context length.
VRAM at each real build (8k context)
DeepSeek V4 Flash 0731 284B ships as a handful of fixed prebuilt files rather than a standard quantization ladder, so it has no real Q4_K_M/FP16/etc. equivalent to DeepSeek V4.1 Flash 552B's. Ranked smallest to largest instead.
| Rank | DeepSeek V4.1 Flash 552B | GB | DeepSeek V4 Flash 0731 284B | GB |
|---|---|---|---|---|
| 1 | Q6_K | 571.6 GB | UD-IQ1_S | 92.5 GB |
| 2 | Q5_K_M | 571.6 GB | UD-Q2_K_XL | 108.5 GB |
| 3 | Q4_K_M | 571.6 GB | UD-IQ3_XXS | 116.6 GB |
| 4 | Q3_K_M | 571.6 GB | UD-Q4_K_XL | 173.8 GB |
| 5 | Q2_K | 571.6 GB | UD-Q8_K_XL | 181.4 GB |
| 6 | NVFP4 | 571.6 GB | — | — |
| 7 | Q8_0 | 657.2 GB | — | — |
| 8 | BF16 | 1236.5 GB | — | — |
| 9 | FP16 | 1236.5 GB | — | — |
| 10 | FP32 | 2473.0 GB | — | — |
Each column is that model's own real builds, smallest first; rank pairs them by position, not by matching precision.
Model specifications
| Spec | DeepSeek V4.1 Flash 552B | DeepSeek V4 Flash 0731 284B |
|---|---|---|
| Org | DeepSeek | DeepSeek |
| Parameters | 552B | 284B |
| Architecture | MoE (16B active) | MoE (13B active) |
| Context | 1024k tokens | 1024k tokens |
| Modalities | text, vision | text |
| License | MIT | MIT |
| Commercial | Yes | Yes |
| Released | 2026-09-10 | 2026-07-31 |
| GPUs (native) | 0 / 119 | 16 / 119 |
Benchmark scores
| Benchmark | DeepSeek V4.1 Flash 552B | DeepSeek V4 Flash 0731 284B |
|---|---|---|
| GPQA Diamond | 90.9 | 91.0 |
| Terminal-Bench 2.1 | 90.6 | 82.7 |
Green = higher score (better). N/A = not yet available. ~ = inherited from a base model, not independently reported for that release itself.
GPUs that run only DeepSeek V4.1 Flash 552B(0)
Every GPU that runs DeepSeek V4.1 Flash 552B also runs DeepSeek V4 Flash 0731 284B.
GPUs that run only DeepSeek V4 Flash 0731 284B(16)
- NVIDIA B300 288GB288 GB
- NVIDIA B200 180GB180 GB
- NVIDIA H200 141GB141 GB
- NVIDIA DGX Spark (128GB)128 GB
- AMD Instinct MI300X192 GB
- AMD Strix Halo (128GB)128 GB
- Apple M5 Ultra (512GB)512 GB
- Apple M5 Ultra (256GB)256 GB
- Apple M5 Max (128GB)128 GB
- Apple M4 Max (128GB)128 GB
- +6 more
Which should you use?
- • You want maximum capability and have a 572 GB+ GPU
- • You need vision/image understanding
- • It's the newer release (2026-09-10 vs 2026-07-31); check the benchmark table above for what actually improved
- • You have limited VRAM: it's a smaller model needing 108.5 GB vs 571.6 GB
Frequently asked questions
- Which is better, DeepSeek V4.1 Flash 552B or DeepSeek V4 Flash 0731 284B?
- DeepSeek V4.1 Flash 552B has 552B parameters vs 284B for DeepSeek V4 Flash 0731 284B, so DeepSeek V4.1 Flash 552B is the larger model. DeepSeek V4 Flash 0731 284B is more hardware-efficient, needing 108.5 GB at its UD-Q2_K_XL build vs 571.6 GB for DeepSeek V4.1 Flash 552B's Q4_K_M. DeepSeek V4 Flash 0731 284B runs on more GPUs natively (16 vs 0).
- How much VRAM does DeepSeek V4.1 Flash 552B need vs DeepSeek V4 Flash 0731 284B?
- At 8k context, DeepSeek V4.1 Flash 552B needs approximately 571.6 GB of VRAM at its Q4_K_M build, while DeepSeek V4 Flash 0731 284B needs 108.5 GB at its UD-Q2_K_XL build. At the largest build each ships, DeepSeek V4.1 Flash 552B requires 1236.5 GB (FP16) vs 181.4 GB (UD-Q8_K_XL) for DeepSeek V4 Flash 0731 284B.
- Can you run DeepSeek V4.1 Flash 552B on the same GPUs as DeepSeek V4 Flash 0731 284B?
- These models have very different VRAM requirements, so they do not share the same compatible GPU set.
- What is the difference between DeepSeek V4.1 Flash 552B and DeepSeek V4 Flash 0731 284B?
- DeepSeek V4.1 Flash 552B has 552B parameters (16B active, MoE) with a 1024k context window. DeepSeek V4 Flash 0731 284B has 284B parameters (13B active, MoE) with a 1024k context window.
- Which model fits in 24 GB of VRAM, DeepSeek V4.1 Flash 552B or DeepSeek V4 Flash 0731 284B?
- Neither fits in 24 GB: DeepSeek V4.1 Flash 552B needs 571.6 GB at Q4_K_M and DeepSeek V4 Flash 0731 284B needs 108.5 GB at UD-Q2_K_XL. Both require a multi-GPU server with 572 GB+ of combined VRAM.
- Which handles long context better, DeepSeek V4.1 Flash 552B or DeepSeek V4 Flash 0731 284B?
- At the full 1,048,576-token window, V4.1 Flash's KV cache is 2.3 GB, against V4 Flash 0731's already-small 9.2 GB, a roughly 4x reduction on top of a design that was already far ahead of a conventional dense-attention model at this context length.