Qwen3.8-Flash-Next vs DeepSeek V4 Flash 284B
Side-by-side VRAM requirements, benchmark scores, and GPU compatibility for local AI inference.
Quick verdict
Qwen3.8-Flash-Next is more hardware-efficient: it needs 123.0 GB at its Q4_K_M build vs 193.8 GB for DeepSeek V4 Flash 284B's Q4_K_M, fitting on 22 GPUs natively.
Analysis
Qwen3.8-Flash-Next and DeepSeek V4 Flash are both preview checkpoints, released four months apart, each showing what its maker's next architecture step looks like before a finished version follows. DeepSeek's preview already has: the identical-architecture DeepSeek V4 Flash 0731 replaced it three months later at no VRAM cost, re-post-trained for agentic work. Qwen's preview is brand new, with its own production successor, 'Qwen3.8-Flash', still to come via Qwen Cloud's API.
DeepSeek V4 Flash and its 0731 successor are spec-identical: 284B total parameters, 13B active, 43 layers, the same hybrid Compressed/Heavily Compressed Attention with a real 10x KV-cache reduction (2.3 GB at Qwen3.8-Flash-Next's own 262,144-token ceiling, against Qwen's 6.4 GB), so everything the 0731 comparison says about weights and cache size against Qwen3.8-Flash-Next applies here unchanged. What's different is what there is to actually run: no community quant repo targets this April checkpoint specifically anymore, Unsloth and nearly everyone else moved straight to quantizing 0731 instead, so the generic quant-ladder math this site computes for it, 108.2 GB at Q2_K, describes a build nobody has actually packaged. Neither model gives the other a benchmark to compare against directly either: DeepSeek reports only MMLU-Pro (86.3) for this checkpoint, a field Qwen3.8-Flash-Next doesn't report at all, while Qwen's GPQA Diamond and LiveCodeBench scores have no DeepSeek V4 Flash figure to sit next to. DeepSeek's own nine-benchmark agentic comparison against its 0731 successor, not against any Qwen model, is the closest either vendor got to grading this specific checkpoint.
Bottom line: There isn't really a live decision here: DeepSeek's own team already answered it by shipping V4 Flash 0731 as a strict upgrade at an identical hardware bill, so nobody should reach for this April checkpoint today when the July rebuild costs nothing extra and adds real agentic training. The comparison is worth reading for the parallel timeline instead: DeepSeek finished its preview-to-flagship cycle in three months, while Qwen3.8-Flash-Next's equivalent cycle, this preview to a finished 'Qwen3.8-Flash', is just getting started.
Total vs. active parameters: two different sparsity ratios
Both models are MoE, so total parameters (the VRAM bill) and active parameters (the compute bill) move independently. DeepSeek routes through a larger slice of a larger model; Qwen routes through a smaller slice of a smaller one.
The same compression, on a checkpoint nobody quantizes anymore
DeepSeek V4 Flash's hybrid attention is architecturally identical to its 0731 successor, so its KV-cache math against Qwen3.8-Flash-Next is the same story, restricted to the 262,144-token window both can reach.
KV cache only, at FP16. Weights and activation overhead sit on top of these figures.
At 262,144 tokens, DeepSeek V4 Flash's real 10x KV compression keeps its cache at 2.3 GB, against Qwen3.8-Flash-Next's 6.4 GB, the identical gap the 0731 successor shows, since the two share the same hybrid attention design.
VRAM at each quantization (8k context)
| Quant | Qwen3.8-Flash-Next | DeepSeek V4 Flash 284B | Diff |
|---|---|---|---|
| FP32 | 806.6 GB | 1272.4 GB | -37% |
| BF16 | 403.4 GB | 636.2 GB | -37% |
| FP16 | 403.4 GB | 636.2 GB | -37% |
| Q8_0 | 214.5 GB | 338.2 GB | -37% |
| Q6_K | 165.7 GB | 261.2 GB | -37% |
| Q5_K_M | 143.8 GB | 226.6 GB | -37% |
| Q4_K_M | 123.0 GB | 193.8 GB | -37% |
| Q3_K_M | 97.2 GB | 153.1 GB | -37% |
| Q2_K | 77.0 GB | 121.3 GB | -36% |
| NVFP4 | 101.0 GB | 159.1 GB | -37% |
Diff is Qwen3.8-Flash-Next relative to DeepSeek V4 Flash 284B. Green = lower VRAM (fits more GPUs).
Model specifications
| Spec | Qwen3.8-Flash-Next | DeepSeek V4 Flash 284B |
|---|---|---|
| Org | Alibaba | DeepSeek |
| Parameters | 180B | 284B |
| Architecture | MoE (6B active) | MoE (13B active) |
| Context | 256k tokens | 1024k tokens |
| Modalities | text, vision, video | text |
| License | Qwen Community 1.0 | MIT |
| Commercial | Yes | Yes |
| Released | 2026-08-26 | 2026-04-24 |
| GPUs (native) | 22 / 119 | 10 / 119 |
Benchmark scores
| Benchmark | Qwen3.8-Flash-Next | DeepSeek V4 Flash 284B |
|---|---|---|
| GPQA Diamond | 91.7 | N/A |
| LiveCodeBench | 91.9 | N/A |
| SWE-bench Pro | 62.5 | N/A |
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 Qwen3.8-Flash-Next(12)
- NVIDIA RTX Pro 600096 GB
- NVIDIA DGX Spark (128GB)128 GB
- AMD Strix Halo (128GB)128 GB
- AMD Strix Halo (96GB)96 GB
- Apple M5 Ultra (96GB)96 GB
- Apple M5 Max (128GB)128 GB
- Apple M4 Max (128GB)128 GB
- Apple M3 Ultra (96GB)96 GB
- Apple M3 Max (128GB)128 GB
- Apple M3 Max (96GB)96 GB
- +2 more
GPUs that run only DeepSeek V4 Flash 284B(0)
Every GPU that runs DeepSeek V4 Flash 284B also runs Qwen3.8-Flash-Next.
GPUs that run both natively(10)
- NVIDIA B300 288GB288 GB
- NVIDIA B200 180GB180 GB
- NVIDIA H200 141GB141 GB
- AMD Instinct MI300X192 GB
- Apple M5 Ultra (512GB)512 GB
- Apple M5 Ultra (256GB)256 GB
- Apple M3 Ultra (512GB)512 GB
- Apple M3 Ultra (256GB)256 GB
- Apple M2 Ultra (192GB)192 GB
- Intel Data Center GPU Max 1550128 GB
Which should you use?
- • You have limited VRAM: it's a smaller model needing 123.0 GB vs 193.8 GB
- • You're running coding tasks
- • You need chain-of-thought reasoning
- • You need vision/image understanding
- • It's the newer release (2026-08-26 vs 2026-04-24); check the benchmark table above for what actually improved
- • You want maximum capability and have a 194 GB+ GPU
- • Long context matters: it supports 1024k tokens vs 256k
Frequently asked questions
- Which is better, Qwen3.8-Flash-Next or DeepSeek V4 Flash 284B?
- Qwen3.8-Flash-Next has 180B parameters vs 284B for DeepSeek V4 Flash 284B, so DeepSeek V4 Flash 284B is the larger model. Qwen3.8-Flash-Next is more hardware-efficient, needing 123.0 GB at its Q4_K_M build vs 193.8 GB for DeepSeek V4 Flash 284B's Q4_K_M. Qwen3.8-Flash-Next runs on more GPUs natively (22 vs 10).
- How much VRAM does Qwen3.8-Flash-Next need vs DeepSeek V4 Flash 284B?
- At 8k context, Qwen3.8-Flash-Next needs approximately 123.0 GB of VRAM at its Q4_K_M build, while DeepSeek V4 Flash 284B needs 193.8 GB at its Q4_K_M build. At the largest build each ships, Qwen3.8-Flash-Next requires 403.4 GB (FP16) vs 636.2 GB (FP16) for DeepSeek V4 Flash 284B.
- Can you run Qwen3.8-Flash-Next on the same GPUs as DeepSeek V4 Flash 284B?
- Yes, 10 GPUs can run both natively in VRAM, including NVIDIA B300 288GB, NVIDIA B200 180GB, NVIDIA H200 141GB. However, 12 GPUs can run Qwen3.8-Flash-Next but not DeepSeek V4 Flash 284B, and no GPU can run DeepSeek V4 Flash 284B without also fitting Qwen3.8-Flash-Next.
- What is the difference between Qwen3.8-Flash-Next and DeepSeek V4 Flash 284B?
- Qwen3.8-Flash-Next has 180B parameters (6B active, MoE) with a 256k context window. DeepSeek V4 Flash 284B has 284B parameters (13B active, MoE) with a 1024k context window. Licensing differs: Qwen3.8-Flash-Next is Qwen Community 1.0 while DeepSeek V4 Flash 284B is MIT.
- Which model fits in 24 GB of VRAM, Qwen3.8-Flash-Next or DeepSeek V4 Flash 284B?
- Neither fits in 24 GB: Qwen3.8-Flash-Next needs 123.0 GB at Q4_K_M and DeepSeek V4 Flash 284B needs 193.8 GB at Q4_K_M. Both require a multi-GPU server with 194 GB+ of combined VRAM.
- Which handles long context better, Qwen3.8-Flash-Next or DeepSeek V4 Flash 284B?
- At 262,144 tokens, DeepSeek V4 Flash's real 10x KV compression keeps its cache at 2.3 GB, against Qwen3.8-Flash-Next's 6.4 GB, the identical gap the 0731 successor shows, since the two share the same hybrid attention design.