DeepSeek V4 Flash 0731: Hardware Requirements, Benchmarks & Self-Hosting Guide
Published: August 3, 2026 · Last updated: August 3, 2026
Quick Navigation
DeepSeek released V4-Flash-0731 on July 31, 2026 — the official, production-ready version of the V4-Flash model that first appeared as a preview in April. It is a Mixture-of-Experts model with 284B total parameters and only 13B active per token, making it dramatically cheaper to run than dense models of similar quality. It also ships with the DSpark speculative decoding module, which boosts per-user generation speed by 60–85%.
For anyone building or buying AI hardware, the key question is simple: what does it take to run this model locally, and is it worth it versus the API? This guide answers both.
All benchmarks are vendor-reported. DeepSeek evaluated V4-Flash-0731 on its own (unreleased) harness at max reasoning effort. No third-party reproductions exist as of August 3, 2026. Treat these numbers as upper bounds until independent labs (Artificial Analysis, LiveBench, Aider leaderboard) publish verified results.
What Is DeepSeek V4 Flash 0731?
The "0731" in the name marks the release date: July 31, 2026. This version supersedes the April V4-Flash Preview with full post-training (re-post-trained from the preview weights), pushing agentic benchmark scores up massively — Terminal Bench 2.1 jumped from 61.8 to 82.7, and DeepSWE rocketed from 7.3 to 54.4.
Key facts at a glance:
The model is MIT licensed — no gating, no commercial restrictions. You can use it in production systems, wrap it in a product, or fine-tune it freely.
Architecture Overview
V4-Flash-0731 inherits the DeepSeek-V4 family architecture with several efficiency-focused design choices:
Mixture-of-Experts
- 1 shared expert + 256 routed experts per MoE layer (intermediate dim 2048)
- 6 experts fire per token out of 256 — this is what keeps active parameters at just 13B despite 284B total
- First 3 MoE layers use hash routing (deterministic, lower overhead)
Hybrid Attention
Two attention mechanisms reduce KV cache size dramatically:
- Compressed Sparse Attention (CSA) — sparse attention pattern that compresses the attention matrix
- Heavily Compressed Attention (HCA) — aggressive compression for long-context positions
Net effect: V4-Flash uses only 10% of DeepSeek-V3.2's KV cache at 1M context. This is what makes the million-token context window practical rather than theoretical.
Manifold-Constrained Hyper-Connections (mHC)
Replaces standard multi-head attention with expansion factor 4 and 20 Sinkhorn-Knopp iterations for routing stability. Multi-token prediction depth is set to 1.
DSpark Speculative Decoding
The model ships with a dedicated draft module (DSpark) that generates 7 speculative tokens per step using greedy draft sampling. This provides 60–85% faster per-user generation at matched aggregate throughput — meaningful for self-hosting scenarios where you want responsive single-user interaction.
Training & Quantization
- Pre-trained on 32T+ tokens using the Muon optimizer
- Quantization-aware trained (QAT): routed experts are stored natively in MXFP4; non-expert tensors in FP8/BF16
- This means the "native" weights are already quantized — you are not losing quality by using the official checkpoints
Efficiency summary: V4-Flash requires only 27% of DeepSeek-V3.2's single-token inference FLOPs. Combined with the 10% KV cache figure, the model is purpose-built for cost-efficient inference.
Benchmark Results
All scores below are vendor-reported by DeepSeek on their own harness at max reasoning effort. V4-Flash-0731 values are bolded; the green column marks the best score in each row.
| Benchmark | V4-Flash-0731 | V4-Flash Preview | V4-Pro Preview | GLM-5.2 | Opus-4.8 |
|---|---|---|---|---|---|
| Terminal Bench 2.1 | 82.7 | 61.8 | 72.1 | 81.0 | 85.0 |
| NL2Repo | 54.2 | 39.4 | 38.5 | 48.9 | 69.7 |
| Cybergym | 76.7 | 38.7 | 52.7 | — | 83.1 |
| DeepSWE | 54.4 | 7.3 | 12.8 | 46.2 | 58.0 |
| Toolathlon-Verified | 70.3 | 49.7 | 55.9 | 59.9 | 76.2 |
| Agents' Last Exam | 25.2 | 15.8 | 16.5 | 23.8 | 25.7 |
| AutomationBench Public | 25.1 | 10.8 | 12.8 | 12.9 | 27.2 |
| DSBench-FullStack † | 68.7 | 37.0 | 41.8 | 61.8 | 71.6 |
| DSBench-Hard † | 59.6 | 25.8 | 31.1 | 54.5 | 71.7 |
† Internal test sets. All scores vendor-reported on DeepSeek's own (unreleased) harness. No third-party reproductions as of Aug 3, 2026.
Key Takeaways from the Numbers
- V4-Flash-0731 crushes the V4-Pro Preview on every agentic benchmark despite being a smaller, cheaper model. This is the headline story.
- DeepSWE improvement is staggering: 7.3 → 54.4 from preview to release. The re-post-training was enormously effective for SWE tasks.
- Opus-4.8 still leads across the board, but at a dramatically higher API price. For cost-sensitive agentic workloads, V4-Flash is the value pick.
- GLM-5.2 is the closest competitor on Terminal Bench (81.0 vs 82.7), but trails significantly on SWE and tool-use benchmarks.
Self-Hosting Hardware Requirements
This is where a hardware comparison site earns its keep. DeepSeek publishes official self-hosting requirements via Unsloth's Dynamic GGUFs. Here is the full table:
| Quantization | Model Size | Min Total Memory (RAM + VRAM) | Notes |
|---|---|---|---|
| 1-bit | ~92 GB | 92 GB | Maximum compression; quality degraded but usable for basic tasks |
| 2-bit | ~102 GB | 102 GB | Significant quality loss; acceptable for draft/QA workflows |
| 3-bit (UD-IQ3_XXS) | ~103 GB | 110–135 GB | Recommended sweet spot — best quality/size tradeoff |
| 4-bit (UD-Q4_K_XL) | ~155 GB | 162 GB | Near-lossless (only 7GB smaller than Q8) |
| 8-bit (UD-Q8_K_XL) | 162 GB | 169 GB | Fully lossless — bit-identical to official weights |
"Total memory" means RAM + VRAM combined. For example, a system with 128GB system RAM and 24GB VRAM GPU has 152GB total — enough for 3-bit or 4-bit quantization with partial GPU offload.
What "Total Memory" Means in Practice
Unlike dense models, MoE models store all expert weights in memory but only activate a subset per token. This means:
- Storage is the barrier, not compute — you need enough memory to hold the full model, even though only 13B parameters fire per token
- CPU inference is viable at lower quantizations — the 13B active parameter count means reasonable token generation speeds even without GPU acceleration
- GPU offload helps but is not mandatory — a high-RAM workstation with no discrete GPU can still run the 3-bit quant at usable speeds
Full-Precision Serving
For production vLLM serving at full precision, DeepSeek recommends a single 4×GB300 node. The official vLLM launch command:
vllm serve deepseek-ai/DeepSeek-V4-Flash-0731 \
--trust-remote-code --kv-cache-dtype fp8 --block-size 256 \
--data-parallel-size 4 --enable-expert-parallel \
--moe-backend deep_gemm_mega_moe \
--attention-config '{"use_fp4_indexer_cache": true}' \
--speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'Recommended sampling: temperature=1.0, top_p=0.95 for agentic tasks, top_p=1.0 for other workloads.
API Pricing Analysis
DeepSeek's official API pricing for V4-Flash-0731:
| Tier | Price (per 1M tokens) | Notes |
|---|---|---|
| Input (cache miss) | $0.14 | Standard prompt processing |
| Input (cache hit) | $0.0028 | 50× cheaper — prompt caching is automatic |
| Output | $0.28 | Token generation |
Concurrency limit: 2,500 requests. API supports Responses API format and is adapted for Codex.
How This Compares
| Model | Input ($/M) | Output ($/M) | vs V4-Flash Output |
|---|---|---|---|
| DeepSeek V4-Flash-0731 | $0.14 | $0.28 | — |
| DeepSeek V4-Pro | ~$0.27 | $0.87 | 3.1× more expensive |
| DeepSeek V3 | $0.27 | $1.10 | 3.9× more expensive |
| GPT-4o | $2.50 | $10.00 | 35.7× more expensive |
| Claude Sonnet 4 | $3.00 | $15.00 | 53.6× more expensive |
| Claude Opus 4 | $15.00 | $75.00 | 267.9× more expensive |
At $0.28/1M output tokens, V4-Flash is one of the cheapest frontier-tier agentic models available. For heavy coding agent workloads (Aider, Cline, custom SWE pipelines), the economics are compelling — a full day of continuous agent usage generating 10M output tokens costs $2.80.
API vs Self-Hosting: Break-Even
At $0.28/1M output tokens, the break-even point for self-hosting is much further out than for pricier models. A $2,000 RTX 5090 workstation running the 3-bit quant would need to generate ~7.1 billion output tokens to break even on output costs alone — roughly 2,000 hours of continuous usage at ~100 tokens/second. Self-hosting V4-Flash makes sense when you need privacy, offline operation, or custom fine-tuning — not primarily for cost savings.
Hardware Recommendations by Tier
Entry-Level Self-Hosting (3-bit, ~103GB)
Target: Run V4-Flash-0731 at acceptable quality with minimal spend.
- Mac Studio M3/M4 Ultra (192GB) — ~$4,000–5,600. Unified memory means the full 103GB model fits with room for context. MLX inference works well. This is the simplest path to running V4-Flash locally.
- Dual-GPU workstation (2× RTX 4090 or 2× RTX 5090) with 128GB+ system RAM. The model partially offloads to GPU for faster token generation. ~$3,500–5,000 total.
- High-RAM desktop (128GB DDR5, no discrete GPU) — CPU-only inference via llama.cpp. Slow (~5–10 t/s) but functional. ~$1,200–1,800.
Near-Lossless Self-Hosting (4-bit, ~155GB)
Target: Quality close to original with reasonable inference speed.
- Mac Studio M4 Ultra (256GB or 512GB) — fits the 4-bit or even 8-bit quant with room for a large KV cache. Best single-device option.
- 4× RTX 4090 workstation (96GB VRAM + 128GB RAM) — significant GPU offload, much faster token generation than CPU-only. ~$6,500+.
- 2× RTX 5090 workstation (64GB VRAM + 128GB RAM) — newer GPUs, faster per-card, but less total VRAM than 4× 4090. ~$5,000+.
Production Serving (Full Precision)
Target: Serve V4-Flash-0731 at scale with speculative decoding.
- Single 4×GB300 node via vLLM — DeepSeek's recommended configuration. Enables full-precision inference with DSpark speculative decoding for maximum throughput and lowest latency.
- Multi-GPU H100/H200 cluster — alternative for organizations with existing NVIDIA infrastructure. Use the vLLM recipe with expert parallelism.
How It Compares
DeepSeek V4-Flash vs V4-Pro
V4-Flash beats V4-Pro Preview on every agentic benchmark while costing roughly one-third as much via API. V4-Pro may still be preferable for non-agentic reasoning tasks once independently benchmarked, but for coding agents, tool use, and SWE workloads, Flash is the clear choice.
DeepSeek V4-Flash vs GLM-5.2
GLM-5.2 is the closest competitor on Terminal Bench 2.1 (81.0 vs 82.7) but trails significantly on SWE benchmarks (DeepSWE: 46.2 vs 54.4) and tool use (Toolathlon: 59.9 vs 70.3). GLM-5.2 has the advantage of broad tool ecosystem compatibility via z.ai's coding plan.
DeepSeek V4-Flash vs Opus-4.8
Opus-4.8 wins every benchmark category but costs 268× more per output token. For budget-constrained teams, V4-Flash delivers roughly 90–97% of Opus-4.8's agentic performance at a fraction of the cost. If you need maximum quality on the hardest tasks (DSBench-Hard, Cybergym), Opus-4.8 is worth it. For everything else, V4-Flash is the smarter economic choice.
Self-Hosting vs API
With V4-Flash's extremely low API pricing, self-hosting is harder to justify on cost grounds alone. The main reasons to self-host:
- Data privacy — proprietary codebases, internal documents
- Offline / air-gapped environments
- Custom fine-tuning — domain-specific adaptations
- Unlimited usage — no per-token costs for extreme-volume workloads (10B+ tokens/day)
Frequently Asked Questions
Can I run DeepSeek V4 Flash 0731 locally?
Yes. The 3-bit quantized version (UD-IQ3_XXS, ~103GB) runs on systems with 110–135GB total memory — achievable with a Mac Studio M3/M4 Ultra at 192GB unified memory, or a dual-GPU workstation with 128GB RAM. For lossless inference, you need 169GB+ (8-bit UD-Q8_K_XL).
How much VRAM do I need?
"Total memory" (RAM + VRAM) matters more than VRAM alone for this model. Minimum: 92GB for 1-bit. Recommended: 110–135GB for 3-bit. Lossless: 169GB+. Full-precision production serving: a 4×GB300 node.
Is V4 Flash better than V4-Pro?
For agentic tasks (coding, tool use, SWE), V4-Flash-0731 beats V4-Pro Preview across every benchmark in DeepSeek's harness. It is also roughly one-third the API price. V4-Pro may retain advantages on pure reasoning once independently tested.
What is DSpark speculative decoding?
DSpark is a draft model attached to V4-Flash that predicts 7 tokens ahead. The main model verifies these predictions in parallel, accepting correct ones — resulting in 60–85% faster per-user generation at the same aggregate throughput.
Is the MIT license really unrestricted?
Yes. MIT license permits commercial use, modification, distribution, and private use. No gating, no attribution clauses beyond the license text.
Affiliate Disclosure: CompareAIHardware.com earns commissions from purchases made through links on this page. This does not affect our editorial content or recommendations.