# DeepSeek V4-Pro

> Source: https://aiwiki.ai/wiki/deepseek_v4_pro
> Updated: 2026-08-01
> Fact-checked: 2026-08-01
> Categories: AI Models, Chinese AI, Large Language Models, Mixture of Experts, Open Source AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "DeepSeek V4-Pro." aiwiki.ai, 1 Aug 2026. https://aiwiki.ai/wiki/deepseek_v4_pro
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**DeepSeek V4-Pro** is the flagship model of the [DeepSeek V4](https://aiwiki.ai/wiki/deepseek_v4) family: a [Mixture of Experts](https://aiwiki.ai/wiki/mixture_of_experts) [large language model](https://aiwiki.ai/wiki/large_language_model) with 1.6 trillion total parameters, 49 billion of them activated per token, and a one-million-token [context window](https://aiwiki.ai/wiki/context_window), built by the Hangzhou lab [DeepSeek](https://aiwiki.ai/wiki/deepseek).[1][8] It shipped on April 24, 2026 as a preview alongside the smaller [DeepSeek V4-Flash](https://aiwiki.ai/wiki/deepseek_v4_flash), and the weights went up on [Hugging Face](https://aiwiki.ai/wiki/hugging_face) the same week under the [MIT License](https://aiwiki.ai/wiki/mit_license).[1][9]

More than three months later it was still the preview. As of August 1, 2026 the `deepseek-v4-pro` endpoint was serving the April build, DeepSeek's change log for July 31 stated that "the official release of DeepSeek-V4-Pro will follow soon" without giving a date, and no new V4-Pro repository had appeared in DeepSeek's Hugging Face organisation.[10][17] In the meantime DeepSeek re-post-trained the smaller Flash model and published nine agent benchmarks on which the 13-billion-active Flash build beats the 49-billion-active Pro preview on every row.[10][18] The flagship is, for now, the older, larger and more expensive of the two models, and its own sibling outscores it on agentic work.

| Field | Value |
|---|---|
| Developer | DeepSeek |
| API model id | `deepseek-v4-pro` |
| Build served | DeepSeek-V4-Pro-Preview (April 24, 2026) |
| Architecture | Mixture of Experts, 1.6T total parameters, 49B activated |
| Layers | 61, plus one multi-token-prediction block |
| Experts | 384 routed plus 1 shared, 6 routed experts per token |
| Context length | 1,048,576 tokens |
| Maximum output | 384K tokens |
| Reasoning modes | Non-Think, Think High (default), Think Max |
| Input modalities | Text only |
| Precision | FP4 MoE experts, FP8 elsewhere |
| Checkpoint | 64 safetensors shards, 864,761,623,612 bytes (about 865 GB) |
| Weights license | MIT |
| Price per 1M tokens | $0.003625 cache-hit input, $0.435 cache-miss input, $0.87 output |
| Concurrency | 500 concurrent requests per account |
| Technical report | arXiv:2606.19348, submitted April 26, 2026 |

The [DeepSeek V4](https://aiwiki.ai/wiki/deepseek_v4) article covers the family as a whole, including the V3 lineage, market reception and the wider industry argument the release started. This page is about the Pro model itself.

## Release and current status

DeepSeek announced the V4 series on April 24, 2026 with the line "1M context is now the default across all official DeepSeek services," describing V4-Pro as offering "performance rivaling the world's top closed-source models" and V4-Flash as "your fast, efficient, and economical choice."[9] Both were explicitly preview builds. The Pro model card still opens with the sentence "We present a preview version of **DeepSeek-V4** series," and that repository was last modified on June 22, 2026.[2][4]

Nothing about that changed through the second quarter or through July. The relevant facts as of August 1, 2026:

| Signal | State on August 1, 2026 |
|---|---|
| API build behind `deepseek-v4-pro` | The April preview. The July 31 change log says "This update only upgrades the DeepSeek-V4-Flash API. The DeepSeek-V4-Pro API and the APP/WEB models are unchanged."[10] |
| Official release date | Not given. "The official release of DeepSeek-V4-Pro will follow soon."[10] |
| Hugging Face | No V4-Pro repository created after June 27, 2026 (the DSpark variant); the main repository last modified June 22, 2026.[4][17] |
| Model card text | Still "a preview version of DeepSeek-V4 series."[2] |
| [OpenAI Responses API](https://aiwiki.ai/wiki/openai_responses_api) support | Not implemented. "We will add support for the `deepseek-v4-pro` model in early August 2026."[11] |
| Reasoning effort mapping | A requested effort of `low` is silently promoted to `high`. "We will update the actual mapped effort of deepseek-v4-pro in early August 2026."[12] |

Two of those rows point at the same window. DeepSeek has committed in writing to changing the Pro effort mapping and to adding Responses API support in early August 2026, which is the closest thing to a schedule the company has published for the model. Neither statement mentions new weights.

The awkward part of that timeline is what happened to Flash. On July 31, 2026 DeepSeek put DeepSeek-V4-Flash-0731 into public beta, published its open weights the same day, and released a nine-benchmark comparison chart whose framing was that the new Flash build now beats V4-Pro-Preview "despite its far smaller activated parameter count."[10][18] The result is a product ladder that currently runs backwards on agent work, with the cheaper model on the newer post-training run and the flagship still on an April checkpoint.

## Architecture

V4-Pro and V4-Flash share one architecture and differ in scale. The three changes the technical report foregrounds are a hybrid attention design pairing Compressed Sparse Attention (CSA) with Heavily Compressed Attention (HCA), Manifold-Constrained Hyper-Connections (mHC) in place of ordinary residual connections, and the [Muon optimizer](https://aiwiki.ai/wiki/muon_optimizer) for the bulk of the parameters.[8] Both are treated in depth at [DeepSeek V4](https://aiwiki.ai/wiki/deepseek_v4); what follows is what the Pro checkpoint specifically declares.

### What the shipped configuration declares

The `config.json` in the Pro repository is the most precise public description of the model, and it differs from the Flash configuration on every dimension that matters.[3]

| Field | V4-Pro | V4-Flash |
|---|---|---|
| `num_hidden_layers` | 61 | 43 |
| `hidden_size` | 7,168 | 4,096 |
| `num_attention_heads` | 128 | 64 |
| `head_dim` | 512 | 512 |
| `num_key_value_heads` | 1 | 1 |
| `n_routed_experts` | 384 | 256 |
| `n_shared_experts` | 1 | 1 |
| `num_experts_per_tok` | 6 | 6 |
| `moe_intermediate_size` | 3,072 | 2,048 |
| `vocab_size` | 129,280 | 129,280 |
| `max_position_embeddings` | 1,048,576 | 1,048,576 |
| `num_nextn_predict_layers` | 1 | 1 |

Positional encoding is YaRN rope scaling by a factor of 16 from an original 65,536-token window up to the full 1,048,576, with `beta_fast` 32 and `beta_slow` 1.[3] Routing uses the `noaux_tc` top-k method with a `sqrtsoftplus` scoring function and a routed scaling factor of 2.5. Two of the training-stability mechanisms described in the technical report are visible as plain fields: `swiglu_limit` is set to 10.0, matching the SwiGLU clamping the paper describes, and `hc_mult` is 4 with `hc_sinkhorn_iters` 20, matching the mHC residual-stream expansion factor and the Sinkhorn-Knopp iteration cap.[3][8] The [DeepSeek Sparse Attention](https://aiwiki.ai/wiki/deepseek_sparse_attention) indexer that CSA depends on is configured with 64 heads, a head dimension of 128 and a top-k of 1,024 blocks.

### The per-layer attention schedule

The configuration also settles exactly how the two attention mechanisms are interleaved, because it ships the schedule as an explicit array. `compress_ratios` has 62 entries, one per layer plus the multi-token-prediction block, and each entry is that layer's compression factor:

```
[128, 128, 4, 128, 4, 128, ... , 4, 128, 4, 0]
```

Reading it off: layers 0 and 1 compress by 128, layers 2 through 60 alternate 4 and 128, and the final entry is 0.[3] A ratio of 128 is HCA, which consolidates every 128 tokens into a single key-value entry and then attends densely over the result. A ratio of 4 is CSA, whose learned compressor folds every 4 tokens into one entry before the sparse indexer picks the top 1,024 blocks to attend over. The trailing 0 is the multi-token-prediction block, which uses no compressed attention at all and falls back to the sliding window, declared separately as 128 tokens.[3][8] So the first two layers and every odd-indexed layer thereafter see a heavily compressed global view, and the alternating CSA layers keep a finer-grained, sparsely selected view. The 128x figure the technical report cites for HCA compression is not an approximation; it is the literal value in the file.

### Efficiency

The point of the arrangement is the cost of a long context. The technical report states that at a one-million-token context DeepSeek-V4-Pro requires only 27 percent of the single-token [inference](https://aiwiki.ai/wiki/inference) FLOPs and 10 percent of the KV cache of [DeepSeek V3.2](https://aiwiki.ai/wiki/deepseek_v3_2).[1][8] That is the claim the whole design rests on, and it is why a 1.6-trillion-parameter model can be priced at all competitively at 1M context. Flash goes further on both counts, which is one reason the two models sit at very different price points despite sharing an architecture.

### Precision and quantisation

V4-Pro ships in what DeepSeek calls FP4 + FP8 Mixed precision: MoE expert parameters in [FP4](https://aiwiki.ai/wiki/fp4), most other parameters in [FP8](https://aiwiki.ai/wiki/fp8).[1] The `quantization_config` block declares FP8 E4M3 weights in 128 by 128 blocks with UE8M0 scales, and `expert_dtype` is set to `fp4`.[3] The base checkpoint, DeepSeek-V4-Pro-Base, is FP8 throughout with no FP4 experts, which is why it is nearly twice the size on disk.[1][5]

The FP4 experts are what DeepSeek ships and expects you to run, not a convenience [quantisation](https://aiwiki.ai/wiki/quantization) of some other reference copy. Running them in FP8 instead is possible, but it means converting the weights first: the bundled instructions are to remove the `"expert_dtype": "fp4"` line from `config.json` and add `--expert-dtype fp8` to `convert.py`, which also takes an expert count and a model-parallel degree. DeepSeek's own example sets that degree to 8, and launches inference with `torchrun --nproc-per-node 8`.[22]

## Variants and repositories

DeepSeek publishes three Pro repositories. None of them is a different model in the sense that a version number would imply.

| Repository | What it is | Created | Last modified | Shards | Storage |
|---|---|---|---|---|---|
| `deepseek-ai/DeepSeek-V4-Pro` | Instruct checkpoint, FP4 experts | 2026-04-22 | 2026-06-22 | 64 | 864,761,623,612 bytes |
| `deepseek-ai/DeepSeek-V4-Pro-Base` | Pre-training checkpoint, FP8 throughout | 2026-04-22 | 2026-04-27 | 64 | 1,606,059,713,686 bytes |
| `deepseek-ai/DeepSeek-V4-Pro-DSpark` | Same instruct checkpoint plus a draft module | 2026-06-27 | 2026-07-04 | 66 | 892,755,974,486 bytes |

The instruct repository had roughly 1.5 million downloads and about 5,360 likes when checked on August 1, 2026. Hugging Face's download counter is a rolling 30-day figure that moves through the day, so these are a snapshot rather than a total.[4][5][6]

DSpark is not a new model, and its card says so in the first sentence: the repository "is not a new model. It is the same checkpoint with an additional speculative decoding module attached."[7] It reports the same 1.6T total and 49B activated figures as the main card. DSpark is one of three draft algorithms in DeepSpec, DeepSeek's MIT-licensed codebase for training and evaluating draft models for [speculative decoding](https://aiwiki.ai/wiki/speculative_decoding), alongside DFlash and [EAGLE-3](https://aiwiki.ai/wiki/eagle_decoding). Under [vLLM](https://aiwiki.ai/wiki/vllm) it is switched on with a single flag, `--speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'`, rather than by loading a separate draft model.[7] Unlike the July Flash build, which folded the module into the shipped checkpoint, the Pro instruct repository does not carry it; DSpark remains a separate download for Pro.[6][18]

### Reading the parameter count

The parameter figure Hugging Face shows beside a repository is derived automatically from a tensor-element count in the model API, not from anything the publisher declares, and for these checkpoints it is easy to misread.[4] For Pro the arithmetic happens to work out, but only if you add the right things together.

| Repository | I8 elements (FP4 experts) | F8_E4M3 elements | BF16 | F32 | Hugging Face total |
|---|---|---|---|---|---|
| `DeepSeek-V4-Pro` | 1,572,763,336,704 | 23,169,335,296 | 2,816,899,328 | 87,776,414 | 1,598,839,674,782 |
| `DeepSeek-V4-Pro-Base` | none | 1,595,932,672,000 | 4,670,257,408 | 185,184,414 | 1,600,790,440,862 |
| `DeepSeek-V4-Pro-DSpark` | 1,623,497,637,888 | 23,952,621,568 | 2,954,820,352 | 90,530,058 | 1,650,497,936,906 |

The instruct model's expert tensors and non-expert FP8 tensors sum to 1,595,932,672,000, which is exactly the FP8 element count of the base checkpoint.[4][5] That is a useful cross-check: it confirms the instruct and base models hold the same parameter set, one with the experts quantised to FP4 and one without, and it puts the total at just under 1.6 trillion, matching DeepSeek's headline figure.

The DSpark repository is where the number stops being a parameter count. It reports 1,650,497,936,906, about 51.7 billion higher, and 28 GB more storage across two extra shards, all of it the attached draft module.[6] Quoting that figure as V4-Pro's parameter count overstates the language model by roughly 3 percent. The same trap bites harder on the Flash side, where two repositories DeepSeek says hold the same checkpoint report 165.3 billion and 304.2 billion because of how quantisation scale tensors are counted. DeepSeek's own figure has not moved from 1.6T total and 49B activated since the April launch, and that is the number to use.[1][7][8]

## Inference configuration

### Reasoning modes

V4-Pro exposes three modes. Non-Think returns a fast answer with no extended reasoning. Think High performs deliberate logical analysis and is the default. Think Max, which requires a special system prompt, pushes reasoning to its fullest extent and is the configuration DeepSeek uses for its headline benchmark numbers.[1] For Think Max, DeepSeek recommends a context window of at least 384K tokens so that long chains of thought are not truncated; the [vLLM](https://aiwiki.ai/wiki/vllm) deployment recipe encodes the same advice as `--max-model-len >= 393216`.[1][14]

In the OpenAI-format API the mode is a `reasoning_effort` value of `low`, `high` or `max`, and in the Anthropic format an `output_config.effort` accepting the same three values; only the Responses API form, `{"reasoning": {"effort": ...}}`, accepts `none`, which disables thinking. The default is `high`.[12]

### The effort mapping quirk

V4-Pro has no genuine low-effort setting. DeepSeek's own table maps the requested effort to an actual effort per model, and for `deepseek-v4-pro` a request for `low` is served at `high`:[12]

| Requested effort | `deepseek-v4-flash` | `deepseek-v4-pro` |
|---|---|---|
| `low` | low | high |
| `high` | high | high |
| `max` | max | max |

The documentation attaches a promise to it: "We will update the actual mapped effort of deepseek-v4-pro in early August 2026."[12] Until then, a caller who selects `low` on Pro to save tokens will not save any, which matters given that the model is verbose even by reasoning-model standards.

### Other API behaviour

Thinking mode ignores `temperature`, `top_p`, `presence_penalty` and `frequency_penalty`; they are accepted without error for client compatibility and then have no effect.[12] Reasoning content has a rule of its own: between two user messages, if the model made no tool call, the intermediate `reasoning_content` does not need to be concatenated back into the context, but if it did make a tool call, that content must be passed back to the API in every subsequent turn.[12] That is what makes interleaved thinking work across an agent loop without polluting ordinary chat history.

For self-hosted use DeepSeek recommends `temperature = 1.0` and `top_p = 1.0`.[1] The repository ships no Jinja chat template. Callers are expected to import `encode_messages` and `parse_message_from_completion_text` from the bundled `encoding` directory, turn OpenAI-format messages into an input string, and parse the completion back out.[1]

## Benchmark record

DeepSeek publishes three separate sets of numbers for V4-Pro, and they are not interchangeable. Base-model figures come from the pre-training checkpoint with few-shot prompting and no reasoning. Instruct figures are labelled by reasoning mode, and the headline comparisons use Think Max. The base checkpoint scores 73.5 on MMLU-Pro and the instruct model in Think Max scores 87.5; they are different artifacts, and quoting one under the other's name is wrong by fourteen points.[1]

### Base model

DeepSeek-V4-Pro-Base against its own predecessor and its smaller sibling, all evaluated with the shot counts shown.[1]

| Benchmark (metric) | Shots | V3.2-Base | V4-Flash-Base | V4-Pro-Base |
|---|---|---|---|---|
| [MMLU](https://aiwiki.ai/wiki/mmlu) (EM) | 5 | 87.8 | 88.7 | 90.1 |
| [MMLU-Pro](https://aiwiki.ai/wiki/mmlu-pro) (EM) | 5 | 65.5 | 68.3 | 73.5 |
| [MMLU-Redux](https://aiwiki.ai/wiki/mmlu_redux) (EM) | 5 | 87.5 | 89.4 | 90.8 |
| [MMMLU](https://aiwiki.ai/wiki/mmmlu) (EM) | 5 | 87.9 | 88.8 | 90.3 |
| AGIEval (EM) | 0 | 80.1 | 82.6 | 83.1 |
| C-Eval (EM) | 5 | 90.4 | 92.1 | 93.1 |
| CMMLU (EM) | 5 | 88.9 | 90.4 | 90.8 |
| MultiLoKo (EM) | 5 | 38.7 | 42.2 | 51.1 |
| [SimpleQA Verified](https://aiwiki.ai/wiki/simpleqa_verified) (EM) | 25 | 28.3 | 30.1 | 55.2 |
| [SuperGPQA](https://aiwiki.ai/wiki/supergpqa) (EM) | 5 | 45.0 | 46.5 | 53.9 |
| FACTS Parametric (EM) | 25 | 27.1 | 33.9 | 62.6 |
| TriviaQA (EM) | 5 | 83.3 | 82.8 | 85.6 |
| BBH (EM) | 3 | 87.6 | 86.9 | 87.5 |
| DROP (F1) | 1 | 88.2 | 88.6 | 88.7 |
| HellaSwag (EM) | 0 | 86.4 | 85.7 | 88.0 |
| WinoGrande (EM) | 0 | 78.9 | 79.5 | 81.5 |
| CLUEWSC (EM) | 5 | 83.5 | 82.2 | 85.2 |
| BigCodeBench (Pass@1) | 3 | 63.9 | 56.8 | 59.2 |
| [HumanEval](https://aiwiki.ai/wiki/humaneval) (Pass@1) | 0 | 62.8 | 69.5 | 76.8 |
| [GSM8K](https://aiwiki.ai/wiki/gsm8k) (EM) | 8 | 91.1 | 90.8 | 92.6 |
| MATH (EM) | 4 | 60.5 | 57.4 | 64.5 |
| MGSM (EM) | 8 | 81.3 | 85.7 | 84.4 |
| CMath (EM) | 3 | 92.6 | 93.6 | 90.9 |
| [LongBench-V2](https://aiwiki.ai/wiki/longbench_v2) (EM) | 1 | 40.2 | 44.7 | 51.5 |

Two of these rows are worth pausing on. The jump on knowledge-recall tests is the largest in the table: SimpleQA Verified nearly doubles from V3.2-Base to V4-Pro-Base, and FACTS Parametric more than doubles. Both measure what the model retained from pre-training rather than how it reasons, which is where a near-doubling of parameter count and a larger token budget would be expected to show. Against that, V4-Pro-Base loses to V3.2-Base on BigCodeBench and to V4-Flash-Base on MGSM and CMath, so the scaling is not uniform.

### Instruct model against frontier competitors

The comparison DeepSeek published at launch, with V4-Pro in Think Max mode.[1]

| Benchmark (metric) | Opus-4.6 Max | GPT-5.4 xHigh | Gemini-3.1-Pro High | K2.6 Thinking | GLM-5.1 Thinking | V4-Pro Max |
|---|---|---|---|---|---|---|
| [MMLU-Pro](https://aiwiki.ai/wiki/mmlu-pro) (EM) | 89.1 | 87.5 | 91.0 | 87.1 | 86.0 | 87.5 |
| [SimpleQA Verified](https://aiwiki.ai/wiki/simpleqa_verified) (Pass@1) | 46.2 | 45.3 | 75.6 | 36.9 | 38.1 | 57.9 |
| Chinese-SimpleQA (Pass@1) | 76.4 | 76.8 | 85.9 | 75.9 | 75.0 | 84.4 |
| [GPQA Diamond](https://aiwiki.ai/wiki/gpqa_diamond) (Pass@1) | 91.3 | 93.0 | 94.3 | 90.5 | 86.2 | 90.1 |
| [HLE](https://aiwiki.ai/wiki/humanity_s_last_exam) (Pass@1) | 40.0 | 39.8 | 44.4 | 36.4 | 34.7 | 37.7 |
| [LiveCodeBench](https://aiwiki.ai/wiki/livecodebench) (Pass@1) | 88.8 | not reported | 91.7 | 89.6 | not reported | 93.5 |
| Codeforces (rating) | not reported | 3168 | 3052 | not reported | not reported | 3206 |
| HMMT 2026 Feb (Pass@1) | 96.2 | 97.7 | 94.7 | 92.7 | 89.4 | 95.2 |
| IMOAnswerBench (Pass@1) | 75.3 | 91.4 | 81.0 | 86.0 | 83.8 | 89.8 |
| Apex (Pass@1) | 34.5 | 54.1 | 60.9 | 24.0 | 11.5 | 38.3 |
| Apex Shortlist (Pass@1) | 85.9 | 78.1 | 89.1 | 75.5 | 72.4 | 90.2 |
| [MRCR](https://aiwiki.ai/wiki/mrcr) 1M (MMR) | 92.9 | not reported | 76.3 | not reported | not reported | 83.5 |
| CorpusQA 1M (ACC) | 71.7 | not reported | 53.8 | not reported | not reported | 62.0 |
| [Terminal-Bench](https://aiwiki.ai/wiki/terminal_bench) 2.0 (Acc) | 65.4 | 75.1 | 68.5 | 66.7 | 63.5 | 67.9 |
| [SWE-bench Verified](https://aiwiki.ai/wiki/swe-bench_verified) (resolved) | 80.8 | not reported | 80.6 | 80.2 | not reported | 80.6 |
| [SWE-bench Pro](https://aiwiki.ai/wiki/swe_bench_pro) (resolved) | 57.3 | 57.7 | 54.2 | 58.6 | 58.4 | 55.4 |
| [SWE-bench Multilingual](https://aiwiki.ai/wiki/swe_bench_multilingual) (resolved) | 77.5 | not reported | not reported | 76.7 | 73.3 | 76.2 |
| [BrowseComp](https://aiwiki.ai/wiki/browsecomp) (Pass@1) | 83.7 | 82.7 | 85.9 | 83.2 | 79.3 | 83.4 |
| HLE with tools (Pass@1) | 53.1 | 52.0 | 51.6 | 54.0 | 50.4 | 48.2 |
| GDPval-AA (Elo) | 1619 | 1674 | 1314 | 1482 | 1535 | 1554 |
| MCPAtlas Public (Pass@1) | 73.8 | 67.2 | 69.2 | 66.6 | 71.8 | 73.6 |
| Toolathlon (Pass@1) | 47.2 | 54.6 | 48.8 | 50.0 | 40.7 | 51.8 |

V4-Pro-Max leads the table on three rows: LiveCodeBench at 93.5, the Codeforces rating at 3206, and Apex Shortlist at 90.2. The Codeforces figure was the highest published for any model at the time, above GPT-5.4 xHigh's 3168 and Gemini 3.1 Pro's 3052. It is roughly tied with the field on [SWE-bench Verified](https://aiwiki.ai/wiki/swe-bench_verified) and on BrowseComp, and it is behind on the two long-context retrieval rows, on Terminal-Bench 2.0, and on Humanity's Last Exam both with and without tools. DeepSeek's own summary in the technical report is that V4 "falls marginally short of GPT-5.4 and Gemini-3.1-Pro, suggesting a developmental trajectory that trails state-of-the-art [frontier models](https://aiwiki.ai/wiki/frontier_models) by approximately 3 to 6 months."[8]

Note that the competitor columns are the versions DeepSeek benchmarked in April 2026. Anthropic, OpenAI, Google, [Moonshot AI](https://aiwiki.ai/wiki/moonshot_ai) and Z.ai have all shipped successors since, including [Claude Opus 4.8](https://aiwiki.ai/wiki/claude_opus_4_8), [Kimi K3](https://aiwiki.ai/wiki/kimi_k3) and [GLM-5.2](https://aiwiki.ai/wiki/glm_5_2), so the table is a snapshot of a moment rather than a current standing.

### What the reasoning modes are worth

The same card publishes each benchmark across all three effort levels, which is the clearest picture of what Think Max is actually buying.[1]

| Benchmark (metric) | Pro Non-Think | Pro High | Pro Max |
|---|---|---|---|
| MMLU-Pro (EM) | 82.9 | 87.1 | 87.5 |
| SimpleQA-Verified (Pass@1) | 45.0 | 46.2 | 57.9 |
| Chinese-SimpleQA (Pass@1) | 75.8 | 77.7 | 84.4 |
| GPQA Diamond (Pass@1) | 72.9 | 89.1 | 90.1 |
| HLE (Pass@1) | 7.7 | 34.5 | 37.7 |
| LiveCodeBench (Pass@1) | 56.8 | 89.8 | 93.5 |
| Codeforces (rating) | not reported | 2919 | 3206 |
| HMMT 2026 Feb (Pass@1) | 31.7 | 94.0 | 95.2 |
| IMOAnswerBench (Pass@1) | 35.3 | 88.0 | 89.8 |
| Apex (Pass@1) | 0.4 | 27.4 | 38.3 |
| Apex Shortlist (Pass@1) | 9.2 | 85.5 | 90.2 |
| MRCR 1M (MMR) | 44.7 | 83.3 | 83.5 |
| CorpusQA 1M (ACC) | 35.6 | 56.5 | 62.0 |
| Terminal-Bench 2.0 (Acc) | 59.1 | 63.3 | 67.9 |
| SWE-bench Verified (resolved) | 73.6 | 79.4 | 80.6 |
| SWE-bench Pro (resolved) | 52.1 | 54.4 | 55.4 |
| SWE-bench Multilingual (resolved) | 69.8 | 74.1 | 76.2 |
| BrowseComp (Pass@1) | not reported | 80.4 | 83.4 |
| HLE with tools (Pass@1) | not reported | 44.7 | 48.2 |
| MCPAtlas (Pass@1) | 69.4 | 74.2 | 73.6 |
| Toolathlon (Pass@1) | 46.3 | 49.0 | 51.8 |

Most of the distance is between Non-Think and High, not between High and Max. On the mathematics and competition rows the gap is enormous: HMMT goes from 31.7 to 94.0 simply by turning thinking on, and Apex from 0.4 to 27.4. Max then adds anywhere from nothing to eleven points on top of High, depending on the task: SimpleQA Verified gains 11.7, Apex 10.9 and LiveCodeBench 3.7, while MMLU-Pro gains 0.4 and MRCR 0.2. One row goes backwards: MCPAtlas is 74.2 at High and 73.6 at Max. Given that Max costs a great deal more output, High is the setting most workloads want, which is presumably why it is the default.

### The July 2026 inversion

On July 31, 2026 DeepSeek published nine agent-oriented scores for the re-post-trained Flash build against the two April preview builds. V4-Pro-Preview loses every row.[10][18]

| Benchmark | V4-Flash-0731 | V4-Flash-Preview | V4-Pro-Preview |
|---|---|---|---|
| Terminal Bench 2.1 | 82.7 | 61.8 | 72.1 |
| NL2Repo | 54.2 | 39.4 | 38.5 |
| Cybergym | 76.7 | 38.7 | 52.7 |
| DeepSWE | 54.4 | 7.3 | 12.8 |
| Toolathlon-Verified | 70.3 | 49.7 | 55.9 |
| Agents' Last Exam | 25.2 | 15.8 | 16.5 |
| AutomationBench Public | 25.1 | 10.8 | 12.8 |
| DSBench-FullStack | 68.7 | 37.0 | 41.8 |
| DSBench-Hard | 59.6 | 25.8 | 31.1 |

Three caveats belong with that table, and DeepSeek supplies two of them itself.

The comparison is against the Pro **preview**, not against an official Pro build, because no official Pro build exists. DeepSeek's chart labels the column "DeepSeek-V4-Pro (Preview)."[18] Whatever the coming official Pro release does to these numbers is unknown, and the whole point of the Flash result is that a post-training run moved a model 47 points on DeepSWE without touching the architecture.

The public code-agent rows were produced in DeepSeek's own scaffold. The footnote reads: "For the Code Agent tasks among the public benchmarks above, DeepSeek-V4-Flash-0731 is evaluated with the minimal mode of DeepSeek Harness (to be released) as the agent framework, using the `max` reasoning effort level with `temperature = 1.0, top_p = 0.95`."[18] Agent benchmark scores depend heavily on the harness driving the model, and the parenthetical "to be released" still stood on August 1, so nobody outside DeepSeek can reproduce these runs.

Two of the nine rows are not public benchmarks at all. The dagger on the last two says "DSBench-FullStack is an internal full-stack development test set; DSBench-Hard is an internal test set of difficult coding-agent problems."[18] Those rows cannot be checked by anyone.

The result still stands after discounting for all three, because an independent measurement points the same way. [Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis), which runs its own harness, scores V4-Pro at 44 on its Intelligence Index v4.1 and the July Flash build at 50, and attributes the gap to the newer post-training run rather than to scale.[16] Where DeepSeek reported Terminal-Bench 2.1 at 82.7 for Flash, Artificial Analysis measured 79, roughly four points lower, which is the size of the harness effect on one directly comparable row.

### Independent measurement

Artificial Analysis places V4-Pro at 44 on Intelligence Index v4.1, an aggregate of nine evaluations. Its published operating figures for the model are an output speed of 62.0 tokens per second, a time to first token of 1.79 seconds, and a cost of $176.34 to run the full index at DeepSeek's first-party rates, consuming about 180 million output tokens in the process.[16] That token count is the number worth carrying: verbosity is a real cost driver on a reasoning model regardless of how low the per-token price is, and it is the mechanism by which a headline rate of $0.87 per million output tokens turns into a three-figure bill for one benchmark sweep.

## Pricing

V4-Pro is billed per token on DeepSeek's first-party API, with three rates.[11]

| Rate | Price per 1M tokens |
|---|---|
| Input, cache hit | $0.003625 |
| Input, cache miss | $0.435 |
| Output | $0.87 |

Cache-hit pricing is automatic: DeepSeek applies [context caching](https://aiwiki.ai/wiki/context_caching) itself rather than asking the caller to declare cache keys or lifetimes. The gap between the two input rates is unusually wide. At $0.003625 against $0.435, a cache hit costs exactly one 120th of a cache miss, a discount of about 99 percent, which on a workload with a large stable prefix is the single biggest lever on the bill.[11]

### How it got here

V4-Pro launched at four times its current price and got cheaper in stages.

| Date | Event |
|---|---|
| April 24, 2026 | Preview launch, at four times the current rates: $3.48 per 1M output tokens and $1.74 per 1M input tokens on a cache miss.[15] |
| April 2026 | A 75 percent introductory discount applied across the V4-Pro price list, with an announced expiry date.[13][15] |
| April 29, 2026 | "The DeepSeek-V4-Pro discount has been extended until May 31, 2026, 15:59 UTC!"[13] |
| May 22, 2026 | "We are making our discount permanent!"[20] The discounted rates become the standing list prices. |

Engadget reported the before and after figures as $0.0145 to $0.003625 on cache-hit input and $3.48 to $0.87 on output, a straight 75 percent cut; the cache-miss input rate moved by the same factor, from $1.74 to $0.435.[15] DeepSeek's pricing page now shows those rates with no promotional label and no expiry attached.[11]

### The peak pricing that has not arrived

A footnote on the pricing page states that DeepSeek "will soon adopt a peak/off-peak pricing policy" under which "during peak hours, prices will be 2x the regular prices, applicable to all billing items," with peak hours defined as 09:00 to 12:00 and 14:00 to 18:00 Beijing time (UTC+8) daily and the effective date left to a future announcement.[11] As of August 1, 2026 the wording was unchanged, no start date had been published, and the July 31 Flash launch came and went without it being switched on. The rates in the table above are what is actually charged. If the policy does take effect, V4-Pro output would run at $1.74 per million tokens during most of a Chinese working day.

For how these rates sit against other vendors, see [LLM API pricing comparison](https://aiwiki.ai/wiki/llm_api_pricing_comparison).

## Availability

### API

`deepseek-v4-pro` is served from `https://api.deepseek.com` for OpenAI ChatCompletions-format calls and from `https://api.deepseek.com/anthropic` for Anthropic Messages-format calls, both since the April launch.[9] The concurrency limit is 500 simultaneous requests, five times tighter than Flash's 2,500. Limits are counted per account rather than per API key, exceeding them returns HTTP 429, and DeepSeek grants capacity expansion on request at no additional cost. A `user_id` parameter (matching `[a-zA-Z0-9\-_]+`, up to 512 characters) provides scheduling isolation between a customer's own end users, and on expanded accounts each `user_id` gets its own concurrency allowance.[19]

The third wire format is where Pro is behind. DeepSeek added the [OpenAI Responses API](https://aiwiki.ai/wiki/openai_responses_api), and with it a documented [OpenAI Codex](https://aiwiki.ai/wiki/openai_codex) configuration, on July 31, 2026 for Flash only. The pricing page, the Responses API guide and the Codex integration page all carry the same restriction, and all three say support for `deepseek-v4-pro` is coming in early August 2026.[11] Until it lands, Codex users pointing at DeepSeek get the smaller model.

Pro is not absent from agent tooling, though. DeepSeek's documented [Claude Code](https://aiwiki.ai/wiki/claude_code) integration through the `/anthropic` endpoint sets `ANTHROPIC_DEFAULT_OPUS_MODEL` and `ANTHROPIC_DEFAULT_SONNET_MODEL` to `deepseek-v4-pro` and leaves the Haiku and subagent roles on `deepseek-v4-flash`, so in that configuration Pro is the model doing the primary work and Flash handles the cheap calls.[23]

### Weights

The instruct and base checkpoints have been public since April 2026 under the MIT License, which permits commercial use, redistribution and modification without royalties.[1][2] Both are on Hugging Face under the `deepseek-ai` organisation. Because the API is serving the same preview build the repository holds, the hosted model and the [open weights](https://aiwiki.ai/wiki/open_weights) are the same artifact, which is not true of every frontier-adjacent release.

What is missing is any weights release for the announced official Pro build. As of August 1, 2026 no such repository existed.[17]

## Hardware requirements

V4-Pro is a difficult model to self-host. The instruct checkpoint is about 865 GB across 64 shards before any KV cache is allocated, and the base checkpoint is roughly 1.61 TB.[4][5]

The vLLM deployment recipe gives concrete configurations.[14] On 8 NVIDIA B300 GPUs it runs single-node with data parallelism plus expert parallelism. On 8 H200s the same layout works but the context has to be capped at 800K rather than the full 1M "to leave KV headroom with dense params replicated across ranks," which is a useful reminder that the advertised context is hardware-dependent in practice. On AMD, the recipe validates 8 MI355X GPUs at 288 GB each under ROCm with tensor parallelism 8, an FP8 KV cache, `--trust-remote-code` and the `deepseek_v4` tokenizer, reasoning and tool-call parsers. On GB200 NVL4 the recipe notes that "the ~960 GB mixed-precision checkpoint does not fit on one tray" and requires multi-node operation across two trays, 8 GPUs in total. An NVFP4 variant needs Blackwell-class hardware for the FP4 indexer cache. The recipe requires vLLM 0.20.0 or newer.

[SGLang](https://aiwiki.ai/wiki/sglang)'s cookbook publishes verified single-node B200, B300 and GB300 configurations and multi-node GB200 configurations, each with low-latency, balanced and high-throughput variants.[21] Two operational warnings there are worth knowing: enabling speculative decoding makes SGLang reset `--max-running-requests` to 48 unless it is set explicitly, which will quietly cap throughput; and context parallelism cannot currently be combined with DP-attention because current releases assert `dp_size == 1` for the interleave layout, so that combination fails at startup.

The practical floor, then, is eight datacentre GPUs with 141 GB or more of memory each, and even that costs you a fifth of the context window. There is no single-node configuration at 1M context on H200-class hardware.

## Limitations

**It is still a preview.** The most consequential limitation is not a capability gap but a release gap. Three months and a week after launch the flagship was still the April build, with no date for the official release, no Responses API support, and a reasoning-effort mapping that DeepSeek itself describes as needing an update.[10][11][12]

**Its own smaller sibling beats it on agent work.** On all nine benchmarks DeepSeek published on July 31, 2026, V4-Flash-0731 scores above V4-Pro-Preview, and Artificial Analysis independently rates Flash six points higher on its aggregate index.[16][18] For agentic coding and tool use as of August 2026, Pro is the wrong model in DeepSeek's own line-up, at three times the price.

**No genuine low-effort mode.** A `low` effort request on Pro is served at `high`, so there is no cheap fast path on the flagship until DeepSeek changes the mapping.[12]

**Text only.** V4-Pro accepts and produces text, nothing else. There is no image or document input at any tier.[1]

**Verbosity.** Artificial Analysis needed roughly 180 million output tokens and $176.34 to run its index on V4-Pro.[16] At Think Max the token cost is high enough to erode a good part of the per-token price advantage.

**Long-context retrieval trails.** On MRCR at 1M tokens Pro scores 83.5 MMR against Opus-4.6 Max's 92.9, and on CorpusQA 1M it scores 62.0 against 71.7.[1] The context window is real and cheap to run, but retrieval quality inside it is not the best available.

**Self-hosting is a datacentre exercise.** About 865 GB of weights, eight high-memory accelerators as the practical floor, and a context capped at 800K on H200-class nodes.[4][14]

**Vendor benchmarks are partly unreproducible.** The July 2026 agent figures came from an unreleased first-party harness, two of the nine benchmarks are internal test sets, and the one directly comparable public row came in about four points lower when measured independently.[16][18]

## See also

- [DeepSeek V4](https://aiwiki.ai/wiki/deepseek_v4)
- [DeepSeek V4-Flash](https://aiwiki.ai/wiki/deepseek_v4_flash)
- [DeepSeek](https://aiwiki.ai/wiki/deepseek)
- [DeepSeek V3.2](https://aiwiki.ai/wiki/deepseek_v3_2)
- [DeepSeek Sparse Attention](https://aiwiki.ai/wiki/deepseek_sparse_attention)
- [Mixture of Experts](https://aiwiki.ai/wiki/mixture_of_experts)
- [Muon (optimizer)](https://aiwiki.ai/wiki/muon_optimizer)
- [Speculative decoding](https://aiwiki.ai/wiki/speculative_decoding)
- [FP4](https://aiwiki.ai/wiki/fp4)
- [FP8](https://aiwiki.ai/wiki/fp8)
- [Quantization](https://aiwiki.ai/wiki/quantization)
- [Open weights](https://aiwiki.ai/wiki/open_weights)
- [MIT License](https://aiwiki.ai/wiki/mit_license)
- [vLLM](https://aiwiki.ai/wiki/vllm)
- [SGLang](https://aiwiki.ai/wiki/sglang)
- [Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis)
- [Context caching](https://aiwiki.ai/wiki/context_caching)
- [LLM API pricing comparison](https://aiwiki.ai/wiki/llm_api_pricing_comparison)
- [Kimi K3](https://aiwiki.ai/wiki/kimi_k3)
- [GLM-5.2](https://aiwiki.ai/wiki/glm_5_2)
- [AI API cost and context planner](https://aiwiki.ai/tools/ai_cost_calculator?model=deepseek-v4-pro)
- [AI model comparison finder](https://aiwiki.ai/tools/compare_models?models=deepseek-v4-pro,deepseek-v4-flash)

## References

1. [deepseek-ai/DeepSeek-V4-Pro model card, Hugging Face (accessed August 1, 2026)](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)
2. [deepseek-ai/DeepSeek-V4-Pro README.md, Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/main/README.md)
3. [deepseek-ai/DeepSeek-V4-Pro config.json, Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/main/config.json)
4. [Model record for deepseek-ai/DeepSeek-V4-Pro, Hugging Face API (accessed August 1, 2026)](https://huggingface.co/api/models/deepseek-ai/DeepSeek-V4-Pro)
5. [Model record for deepseek-ai/DeepSeek-V4-Pro-Base, Hugging Face API (accessed August 1, 2026)](https://huggingface.co/api/models/deepseek-ai/DeepSeek-V4-Pro-Base)
6. [Model record for deepseek-ai/DeepSeek-V4-Pro-DSpark, Hugging Face API (accessed August 1, 2026)](https://huggingface.co/api/models/deepseek-ai/DeepSeek-V4-Pro-DSpark)
7. [deepseek-ai/DeepSeek-V4-Pro-DSpark model card, Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-DSpark)
8. [DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence, arXiv:2606.19348 (submitted April 26, 2026)](https://arxiv.org/abs/2606.19348)
9. [DeepSeek-V4 Preview Release, DeepSeek API Docs (April 24, 2026)](https://api-docs.deepseek.com/news/news260424)
10. [Change Log, DeepSeek API Docs (entry dated 2026-07-31, accessed August 1, 2026)](https://api-docs.deepseek.com/updates)
11. [Models & Pricing, DeepSeek API Docs (accessed August 1, 2026)](https://api-docs.deepseek.com/quick_start/pricing)
12. [Thinking Mode, DeepSeek API Docs (accessed August 1, 2026)](https://api-docs.deepseek.com/guides/thinking_mode)
13. [The DeepSeek-V4-Pro discount has been extended until May 31, 2026, @deepseek_ai on X (April 29, 2026)](https://x.com/deepseek_ai/status/2049312932014813344)
14. [DeepSeek-V4-Pro deployment recipe, vLLM Recipes (accessed August 1, 2026)](https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4-Pro)
15. [DeepSeek permanently reduces the price of its flagship V4 model by 75 percent, Engadget (May 23, 2026)](https://www.engadget.com/2180062/deepseek-permanently-reduces-the-price-of-its-flagship-v4-model-by-75-percent/)
16. [DeepSeek V4 Pro: Intelligence, Performance & Price Analysis, Artificial Analysis (accessed August 1, 2026)](https://artificialanalysis.ai/models/deepseek-v4-pro)
17. [deepseek-ai models sorted by creation date, Hugging Face API (accessed August 1, 2026)](https://huggingface.co/api/models?author=deepseek-ai&sort=createdAt&direction=-1&limit=20)
18. [deepseek-ai/DeepSeek-V4-Flash-0731 model card, Hugging Face (accessed August 1, 2026)](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731)
19. [Rate Limit & Isolation, DeepSeek API Docs (accessed August 1, 2026)](https://api-docs.deepseek.com/quick_start/rate_limit)
20. [We are making our discount permanent, @deepseek_ai on X (May 22, 2026)](https://x.com/deepseek_ai/status/2057854261699195173)
21. [DeepSeek-V4 cookbook, SGLang documentation (accessed August 1, 2026)](https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4)
22. [deepseek-ai/DeepSeek-V4-Pro inference/README.md, Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/main/inference/README.md)
23. [Integrate with Claude Code, DeepSeek API Docs (accessed August 1, 2026)](https://api-docs.deepseek.com/quick_start/agent_integrations/claude_code)