DeepSeek V4 GA: Architecture, Inference Efficiency, and What the Grayscale Test Reveals

A technical analysis of the CSA + HCA hybrid attention mechanism, benchmark performance across reasoning budgets, and the open-weight deployment story — as the 1.6T-parameter MoE model approaches General Availability.

---

The DeepSeek V4 API started behaving strangely around July 17, 2026. Users on identical deepseek-v4-pro model IDs began receiving noticeably sharper outputs. More concise. Fewer hallucinated tool calls. Same endpoint, same parameter string, better results.

By July 18, a rumor swept across X: that DeepSeek was secretly routing API traffic to Anthropic's Claude Fable 5 to collect distillation data. The claim went viral. It is false. The evidence cited — output similarity and CoT style shifts — has straightforward explanations: convergent behavior from models trained on overlapping corpora, and a new instruction-tuning distribution from an updated checkpoint. No technical proof of API-level proxying has been presented. The grayscale test outputs reflect genuine V4 GA progress, not a third-party model behind a proxy.

This article provides a technical analysis of the V4 architecture, the benchmark profile, and the deployment story. We focus on what is independently verifiable: the weights on HuggingFace, the published technical report, and the reproducibility of benchmark scores through open-source inference engines.

---

1. Architecture: The CSA + HCA Hybrid Attention Mechanism

1.1 Model Configuration

DeepSeek V4 is a Mixture-of-Experts Transformer with two variants released under the MIT license. The full specification is available in the technical report and the HuggingFace model cards.

1.2 Attention Design

The primary architectural innovation in V4 is a dual-path hybrid attention mechanism that replaces V3.2's Multi-Head Latent Attention (MLA). The motivation is straightforward: standard attention scales quadratically with sequence length, making 1M-token contexts computationally prohibitive without aggressive compression.

Compressed Sparse Attention (CSA):

• Token-wise compression at a 4:1 ratio along the sequence axis, implemented via a softmax-gated pooling function with learned positional bias
• A "Lightning Indexer" operating in FP4 precision performs ReLU-scored multi-head dot product to select the top-512 compressed blocks per query token
• Inherits sparse-selection routing from V3.2's DeepSeek Sparse Attention (DSA), but operates on blocks compressed 4× shorter, reducing the indexer cost proportionally
• Sliding-window branch over the last 128 uncompressed tokens ensures recency is preserved

Heavily Compressed Attention (HCA):

• 128:1 compression ratio — the sequence is compressed to a length where dense attention is computationally negligible
• No sparse selection: every query attends densely to every compressed block
• Provides global coherence across the full context window without the indexer overhead

Layer Layout (V4-Pro, 61 transformer layers):

Layers 0–1 use HCA exclusively. Layers 2–60 alternate CSA and HCA layers, with each attention type handling distinct positional ranges. A final Multi-Token Prediction (MTP) block uses only sliding-window attention for speculative decoding.

KV-Cache Efficiency (measured at 1M-token context):

KV entries are stored in FP8 across both attention paths, with BF16 reserved exclusively for Rotary Position Embedding (RoPE) dimensions. The 98% KV-cache reduction relative to standard grouped-query attention is the mechanism that makes 1M-token inference economically practical at V4's API price point.

1.3 Training Improvements

Manifold-Constrained Hyper-Connections (mHC): Residual connections are replaced with per-token mixtures over multiple parallel branches, each projected onto a learned manifold via Sinkhorn normalization before recombination. This geometric constraint stabilizes signal propagation across the 1.6T-parameter graph and reduces training instability at scale. The technique is described in the technical report and follows the lineage of hyper-connection research from DeepSeek's V3.2 architecture work.

Muon Optimizer: V4 adopts Muon — matrix momentum with Newton-Schulz orthogonalization — replacing AdamW. The reported benefits include faster convergence (measured in training steps to equivalent validation loss) and reduced sensitivity to learning rate schedule hyperparameters. For researchers fine-tuning from V4 base weights, this is operationally significant: the checkpoint optimization trajectory differs fundamentally from V3-series checkpoints trained under AdamW.

Two-Stage Post-Training Paradigm:

1. Domain-specific expert models are trained independently via Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO) across reasoning, coding, mathematics, world knowledge, and agentic tool-use domains.
2. These distinct proficiencies are consolidated into a single model through on-policy distillation — the consolidated model learns by observing the behavior distribution of domain experts, not merely their output distribution.

This approach addresses the multi-task interference problem common in single-stage reinforcement learning pipelines, where improvements in one domain (e.g., code generation) degrade performance in another (e.g., mathematical reasoning).

---

2. Benchmark Analysis

2.1 Coding Benchmarks

V4-Pro-Max establishes unambiguous open-source leadership on coding benchmarks. The following table compares against the best available closed-source scores.

V4-Pro-Max leads all models — open and closed — on LiveCodeBench, Codeforces Rating, and Apex Shortlist. On SWE-bench Verified, it is in a statistical three-way tie with Opus 4.6 Max (80.8) and GPT-5.4 (80.6). A Codeforces rating of 3206 places the model in the International Grandmaster range — the highest reported rating among all models in the comparison set.

Methodological note on SWE-bench: The SWE-bench Verified verifier has been shown to accept approximately 8.5% of functionally incorrect solutions. On DeepSWE, a contamination-free benchmark constructed across 91 repositories and 5 programming languages, V4-Pro scores 8% pass@1 compared to GPT-5.5 at 70% and Opus 4.7 at 54%. Researchers should calibrate coding evaluations against their own task distribution rather than relying solely on SWE-bench scores.

2.2 Reasoning and Knowledge

The pattern is consistent across knowledge-intensive benchmarks: V4-Pro-Max operates within 3–5 percentage points of the closed-source frontier on reasoning tasks (GPQA Diamond, MMLU-Pro) but trails more significantly on factual precision (SimpleQA) and hardest-reasoning (HLE). The SimpleQA gap — 57.9 vs. 75.6 — is the largest capability delta and suggests that knowledge-intensive production pipelines should pair V4 with retrieval-augmented generation. DeepSeek's own technical report assesses that V4 trails state-of-the-art frontier models by approximately 3 to 6 months in knowledge capability.

2.3 Agentic and Tool-Use Benchmarks

Agentic tasks represent the largest systematic gap between V4-Pro and the closed-source frontier. On MCPAtlas (tool selection and composition) and SWE Multilingual, V4-Pro is effectively tied with the best closed models. On Terminal Bench — which measures autonomous terminal command execution — there remains a 7.2-point gap, consistent with the finding that V4's architecture optimizes for inference efficiency rather than long-horizon agent scaffolding.

2.4 Reasoning Budget Scaling

V4 supports configurable reasoning budgets via the reasoning_effort parameter, which controls the depth of chain-of-thought generation. The scaling behavior is well-characterized across the published benchmarks.

A key empirical finding: V4-Flash at Max reasoning budget (91.6 LiveCodeBench) outperforms V4-Pro at High budget (89.8). For cost-constrained deployments, Flash-Max is the price-performance optimum on reasoning tasks. Pro-Max should be reserved for workloads where the benchmark delta demonstrably affects downstream outcomes.

---

3. Deployment and Inference Economics

3.1 Weights on HuggingFace

All checkpoints are available under the DeepSeek V4 collection. The instruct models use mixed FP4/FP8 precision — MoE expert parameters are stored in FP4, and most other parameters use FP8.

DSpark speculative decoding variants: V4-Flash-DSpark (165B disk, 272K downloads) and V4-Pro-DSpark (889B disk, 43.8K downloads).

3.2 Serving Infrastructure

# vLLM (recommended, requires v0.8.0+)
pip install vllm
vllm serve "deepseek-ai/DeepSeek-V4-Pro" --max-model-len 131072

# SGLang (requires v0.4.0+)
python3 -m sglang.launch_server \
    --model-path "deepseek-ai/DeepSeek-V4-Pro" \
    --host 0.0.0.0 --port 30000

3.3 Hardware Requirements

V4-Flash at INT4 quantization is the first frontier-competitive open model deployable on consumer-adjacent hardware — a significant milestone for the open-source inference ecosystem.

3.4 API Pricing with Peak/Off-Peak Tiers

Effective with GA, DeepSeek introduces time-of-day pricing (USD per 1M tokens):

Peak hours: 09:00–12:00 and 14:00–18:00 Beijing Time (UTC+8), seven days per week.

Context caching is automatic. A cache hit on V4-Pro costs 120× less than a cache miss ($0.003625 vs. $0.435 per 1M input tokens). In agentic loops where identical system prompts and tool definitions are reused across turns, this mechanism is the primary cost driver. Compared to Claude Opus 4.7 at $25/M output, V4-Pro at peak is approximately 14× cheaper; at off-peak, 29× cheaper.

---

4. Grayscale Test and GA Timeline

The grayscale (灰度测试) rollout is a standard software engineering practice: a progressive deployment of the GA-candidate checkpoint to a subset of API users before full release. The signals we observe are consistent with this interpretation:

• July 4: Build IDs deepseek-v4-pro-202606 and deepseek-v4-flash-202605 were observed on X, following DeepSeek's YYYYMM naming convention for internal checkpoints
• July 9: @teortaxesTex reported that V4-Pro API calls had "a chance of routing to the official version"
• July 17: Multiple Bilibili and X users began posting grayscale test outputs demonstrating improved code generation and 3D game creation
• July 18: The grayscale test wave broadened significantly; @ChinaMacroFacts reported that the GA build is "likely stronger than Kimi K3"
• July 19 (current): GA launch expected imminently — community consensus points to July 20 or within the week, potentially timed with WAIC in Shanghai

The July 24, 2026, 15:59 UTC deadline for legacy endpoint retirement (deepseek-chat, deepseek-reasoner) remains the operational forcing function. Post-deadline, all API traffic must use deepseek-v4-flash or deepseek-v4-pro model identifiers.

---

5. Open Weights and Ecosystem Impact

The open-weight release of V4 under MIT license has several implications for the research community:

Verifiability: Unlike API-only models, V4's benchmark scores are independently reproducible by any researcher with sufficient GPU capacity. The grayscale test API outputs — whatever their provenance — do not affect the verifiability of the published weights. Researchers should deploy V4 checkpoints on their own infrastructure for evaluation, not benchmark against API calls during the rollout period.

Inference efficiency as a research contribution: The CSA + HCA hybrid attention mechanism is a publishable finding independent of the model's benchmark scores. A 98% KV-cache reduction at 1M-token context, achieved through a specific combination of compressed sparse and heavily compressed attention layers, advances the state of long-context inference engineering. The approach is likely to be replicated across the open-source ecosystem.

Economic accessibility: V4-Flash at $0.14/M input tokens (cache miss) and $0.28/M output tokens makes frontier-adjacent inference available to research labs and individual developers who could not previously afford models in this capability tier. Combined with INT4 quantization viability on consumer GPUs, V4 lowers the barrier to entry for long-context model research.

The competitive landscape: With Kimi K3 (2.8T MoE), GLM 5.2 (744B MoE), Qwen 3.8 (2.4T, upcoming), and MiniMax M3 Pro (3T, upcoming) all operating in the same parameter scale and MIT-license regime, the open-source frontier model ecosystem has never been more competitive. V4's primary differentiation is inference efficiency — not raw benchmark scores — which positions it as the cost leader in the open-weight category.

---

6. FAQ

Q: Is V4 officially GA yet?

No. As of July 19, 2026, the model is in a grayscale testing phase. The official GA launch is expected within days. The legacy model IDs (deepseek-chat, deepseek-reasoner) retire on July 24, 2026, at 15:59 UTC.

Q: Was the grayscale test actually routing to another model?

No. This is an unsubstantiated community rumor. The evidence — output similarity under specific prompt formulations and CoT-style shifts — is consistent with a genuine checkpoint improvement using a different instruction-tuning distribution. Similar outputs across large models trained on overlapping corpora with convergent post-training objectives are expected, not anomalous. CoT style changes are standard across checkpoint iterations. No technical evidence of API-level proxying has been presented. The grayscale test outputs reflect real V4 GA progress.

Q: How do I verify V4's performance?

Deploy the open weights on your own infrastructure using vLLM or SGLang. Benchmark against your task distribution. The Preview weights on HuggingFace are the known-good baseline. When GA weights are released, re-benchmark and measure the delta. Do not benchmark against API outputs during the grayscale rollout period.

Q: What hardware do I need to self-host V4?

V4-Flash: 2×H100-80GB or 4×A100-80GB for non-thinking mode. V4-Flash INT4 quantized: 2×RTX 6000 Ada (48GB each) at reduced context. V4-Pro: 8×H100-80GB minimum for non-thinking mode; 16×H100-80GB for thinking-max workloads.

Q: What is the legacy endpoint retirement strategy?

deepseek-chat and deepseek-reasoner return HTTP 400 errors after July 24, 2026, 15:59 UTC. Replace with deepseek-v4-flash (non-thinking default) or deepseek-v4-pro (premium tier). Thinking mode is controlled by the thinking parameter, not the model identifier. API keys, base URLs, and billing remain unchanged.

Q: Should I wait for the GA checkpoint before evaluating V4?

No. The Preview weights on HuggingFace represent the known-good checkpoint. Deploy, benchmark, and establish your baseline now. The GA release may bring quality improvements, but the Preview is a valid evaluation target. Use the GA build as an upgrade, not a starting point.

---

7. What Researchers Should Do Now

1. Deploy V4-Flash from open weights and benchmark against your task distribution. vllm serve "deepseek-ai/DeepSeek-V4-Flash" --max-model-len 131072. The Preview weights are the only independently verifiable artifact currently available.

2. Re-benchmark when GA weights are released to HuggingFace. The delta between Preview and GA checkpoint performance on your specific workload tells you how much of the grayscale-test buzz reflects real improvement.

3. Study the CSA + HCA attention mechanism. The hybrid attention design — alternating 4:1 compressed sparse attention with 128:1 heavily compressed attention across layers — is the architecture's primary research contribution. It is publishable, replicable, and likely to influence the next generation of open-source long-context models.

4. Migrate legacy model identifiers before July 24. grep -r "deepseek-chat\|deepseek-reasoner" . — fix every occurrence. The deadline is firm. There is no announced extension.

The V4 architecture is described in the published technical report. The weights are on HuggingFace under the MIT license. The benchmarks are reproducible through standard inference engines. The grayscale test is a deployment mechanism — not a mystery. Deploy the weights, benchmark your workload, and measure what you can verify.

---

Browse the DeepSeek V4 collection on HuggingFace. Deploy with vLLM. Benchmark on your own data. The weights are the ground truth.

For more DeepSeek V4 GA coverage, see Deepseek v4 ga news.