DeepSeek-V4 vs V3.2: 73% Less Compute, Better Performance

On April 24, 2026, DeepSeek officially released the preview version of DeepSeek-V4 and open‑sourced it simultaneously. Barely a few months after the release of V3.2, DeepSeek once again shook the AI community with a series of eye‑catching technological breakthroughs. From V3.2 to V4, this is not just a routine version bump – it marks a paradigm shift from a "parameter race" to an "efficiency revolution". This article provides a comprehensive comparison between these two generations across multiple dimensions: model scale, technical architecture, performance, long‑context capabilities, and cost/pricing.

1. Model Scale: Bigger, but Smarter

DeepSeek‑V3.2 adopts a Mixture of Experts (MoE) architecture with a total of 685 billion parameters, activating about 37 billion per token. The model is built on 61 Transformer layers, has 128 attention heads, a hidden dimension of 7168, and uses 256 routing experts with 8 experts activated per token. As a critical technical validation step towards V4, V3.2 introduced the DeepSeek Sparse Attention (DSA) mechanism for the first time.

DeepSeek‑V4 launches two models at once:

• DeepSeek‑V4‑Pro: 1.6 trillion total parameters (1.6T), with 49 billion activated per token. This makes V4‑Pro the largest open‑weight model known at the time of release, surpassing Kimi K2.6 (1.1T), GLM‑5.1 (754B), and V3.2 (685B).

• DeepSeek‑V4‑Flash: 284 billion total parameters, activating only 13 billion per token, targeting high cost‑effectiveness and low latency.

In terms of parameter count, V4‑Pro has more than 2.3 times the total parameters of V3.2, and its activated parameters have increased from 37B to 49B. However, the growth in parameters is not V4’s most essential evolution – what truly stuns is that V4 drastically reduces computational resource requirements while significantly increasing parameter size.

2. Technical Architecture: A Generational Revolution in Attention Mechanisms

If the core innovation of V3.2 was the DSA (DeepSeek Sparse Attention) mechanism, then V4 completes a thorough architectural overhaul on that foundation.

V3.2’s DSA mechanism: V3.2 implemented fine‑grained sparse attention for the first time, using a dual‑component design of "Lightning Indexer + Fine‑grained Token Selection", establishing a sparse computation paradigm that "filters first, computes later". This mechanism reduces the main attention from quadratic complexity to near‑linear (O(Lk)), boosting long‑text inference speed by 2‑3× and cutting memory usage by 30‑40%.

V4’s hybrid attention architecture: V4 goes further by introducing a novel hybrid of CSA (Compressed Sparse Attention) and HCA (Heavily Compressed Attention) .

• CSA (Compressed Sparse Attention): Compresses KV entries by a factor of 4 along the sequence dimension, and uses the Lightning Indexer to select top‑k compressed blocks for each query. This is akin to "only remembering key points and skimming".

• HCA (Heavily Compressed Attention): Compresses KV entries by a factor of 128, and every query attends densely to each compressed block. This performs another round of deep compression on relationships between distant segments.

In V4‑Pro’s 61‑layer structure, different layers alternately use CSA and HCA. This "layered division of labour" avoids the inefficiency of forcing one mechanism to cover all scenarios.

Two additional innovations are introduced in V4: manifold‑constrained hyperconnection (mHC) replaces traditional residual connections, solving the problem of information decay in deep networks; and the Muon optimizer replaces AdamW, adapted to the MoE structure and low‑precision training.

From an architectural evolution perspective, V3.2’s DSA serves as the technical prelude to V4’s hybrid attention – V4 inherits the sparse‑selection idea of DSA but runs it on sequences already compressed by a factor of 4, further amplifying efficiency gains.

3. Performance: A Comprehensive Qualitative Leap

3.1 Overall Intelligence Index

On the authoritative Artificial Analysis open‑source model leaderboard, V4‑Pro scored 52, a leap of 10 points from V3.2’s 42. V4‑Pro became the second‑largest open‑source reasoning model globally, right after Kimi K2.6.

3.2 Knowledge Capabilities

V4 achieves significant improvements in knowledge‑intensive tasks:

On MMLU‑Pro (EM), V4‑Pro reaches 87.5, matching GPT‑5.4. In world‑knowledge evaluations, V4‑Pro significantly outperforms other open‑source models and trails only Gemini‑Pro‑3.1.

3.3 Reasoning and Coding

V3.2 had already reached GPT‑5 level on reasoning benchmarks, approaching Gemini‑3.0‑Pro. V4 pushes this level even further:

• LiveCodeBench: V4‑Pro scores 93.5, surpassing GPT‑5.4.
• Codeforces Rating: V4‑Pro reaches 3206.
• Apex Shortlist: V4‑Pro scores 90.2.
• HumanEval: V4‑Pro scores 76.8, a 14‑point gain from V3.2’s 62.8.

V4 outperforms Kimi K2.6 Thinking and GLM‑5.1 Thinking in competitive‑level coding evaluations. On the Vals AI Vibe Code Benchmark, V4 ranks first among open‑weight models with an "overwhelming lead", delivering roughly a 10× performance improvement over V3.2.

3.4 Agent Capabilities

This is one of the most significant areas of improvement from V3.2 to V4. V4‑Pro achieves the best performance among open‑source models in Agentic Coding evaluations. In real‑world tests, V4‑Pro can autonomously code for over 60 consecutive minutes without human intervention, completing complex database design and Android emulator development tasks. In terms of long‑horizon planning, self‑correction, and tool calling, V4‑Pro’s performance "far exceeds DeepSeek‑V3.2".

Internal feedback indicates that V4’s user experience surpasses Sonnet 4.5 and its delivery quality is close to Opus 4.6 in non‑thinking mode. DeepSeek states that V4 performs closest to Claude Sonnet in agent and coding tasks and has already been adopted as the primary model for Agentic Coding internally.

4. Long‑Context Capabilities: From 160K to 1M

V3.2 supports a 160K context, which was already considerable at the time. V4 pushes this to 1 million tokens (1M) – both V4 models natively support this ultra‑long context. 1M context is now the standard across all official DeepSeek services.

However, window expansion is only half the story – the truly critical part is cost reduction. Under the 1M context setting:

In other words, V4‑Pro uses only 27% of the compute and 10% of the KV cache of V3.2 to process a context of the same length; V4‑Flash needs just 10% compute and 7% cache. In multiplier terms, V4‑Pro delivers a 3.7× reduction in per‑token compute, and V4‑Flash achieves a 9.8× reduction.

If V3.2’s DSA reduced attention from quadratic to near‑linear complexity, then V4’s hybrid CSA+HCA architecture performs yet another order‑of‑magnitude compression on top of that. DeepSeek attributes this to "a brand‑new attention mechanism design" – but in reality, it is a full‑stack system overhaul covering model structure, training stability, inference caching, and MoE communication.

5. Cost and Pricing: The Commercial Realisation of Efficiency

V3.2’s API pricing (in RMB) was:

• Input (cache hit): 0.2 RMB / 1M tokens
• Input (cache miss): 2 RMB / 1M tokens
• Output: 3 RMB / 1M tokens
• Context: 128K

Thanks to DSA’s efficiency gains, V3.2 already reduced API costs by over 50% compared to its predecessor.

V4’s pricing strategy is more differentiated. Using the exchange rate 1 USD = 6.7 RMB, the converted prices are:

(RMB original: V3.2 – hit 0.2, miss 2, out 3; V4‑Flash – hit 0.2, miss 1, out 2; V4‑Pro – hit 1, miss 12, out 24)

V4‑Flash sees a price cut across the board – its output is only 0.30 USD per 1M tokens, making it one of the cheapest in the small‑model segment. V4‑Pro, on the other hand, commands a premium – its output is 8× that of V3.2. However, considering that V4‑Pro can drastically reduce costs in long tasks through caching, and its performance now rivals top closed‑source models, the price remains competitive: V4‑Pro’s output of 3.58 USD is lower than Gemini 3.1 Pro (12 USD), GPT‑5.4 (15 USD), and Claude Sonnet 4.6 (15 USD).

DeepSeek emphasises that its low prices stem from extreme efficiency optimisation for million‑token scenarios. V4‑Pro’s per‑token compute is only 27% of V3.2’s, meaning the same hardware can handle far more requests.

6. Conclusion and Outlook

From DeepSeek‑V3.2 to DeepSeek‑V4, we have witnessed a profound paradigm shift. V3.2 used DSA sparse attention to elevate long‑text efficiency while preserving quality; V4, through an entire suite of architectural innovations – CSA, HCA, mHC, Muon, and more – delivers a dramatic efficiency leap despite a significant increase in parameter size, slashing per‑token compute to a quarter or even a tenth of its predecessor.

With only 27% of the compute, V4 achieves a quantum leap in inference efficiency, epitomising the core evolutionary logic of Chinese AI moving from "parameter stacking" to "efficiency first". DeepSeek’s official positioning of V4 is clear: V4‑Pro approaches closed‑source systems like Gemini in knowledge and reasoning, but still lags by about 3‑6 months; V4‑Pro‑Max surpasses GPT‑5.2 and Gemini‑3.0‑Pro in reasoning, but trails GPT‑5.4 and Gemini‑3.1‑Pro by a similar margin.

Another significant implication of V4 is its adaptation to domestic AI accelerators. V4 has completed fine‑grained expert parallelism verification on Huawei’s Ascend NPU platform, achieving a speedup of 1.50‑1.73× on general inference workloads, making it the world’s first trillion‑parameter model trained and inferred on a domestically produced AI accelerator.

From V3.2 to V4, DeepSeek has accomplished a pivotal transition from "chasing" to "leading" in under a year. V4 not only tops the open‑source leaderboard but also matches or even surpasses top closed‑source models in multiple core capabilities. The driving force behind all this is not mere parameter stacking, but a profound efficiency revolution – doing more with less.