DeepSeek-V4-Pro-0813 Benchmarks: How It Stacks Up Against Opus and Kimi K3

DeepSeek AI released DeepSeek-V4-Pro-0813, the official non-preview version of its open-weight coding and agentic model, reporting gains over the preview on every listed benchmark. On Terminal Bench 2.1 it scores 87.9, behind Kimi K3 (88.3) and Fable-5 (88.0) but ahead of Opus-4.8 (85.0) and GLM-5.2 (81.0); on Humanity's Last Exam it scores 42.7 without tools and 60.0 with tools, surpassing Opus-4.8's 57.9 tool-assisted score. The model ships with DSpark speculative decoding and an MIT license, and its DeepSWE score jumps from 12.8 to 62.7 versus the preview.

DeepSeek-V4-Pro-0813 Benchmarks: How It Stacks Up Against Opus and Kimi K3 DeepSeek-V4-Pro-0813 benchmark scores across Terminal Bench, HLE, and Cybergym, compared against GLM-5.2, Kimi K3, Opus-4.8, and Fable-5. What is DeepSeek-V4-Pro-0813? DeepSeek-V4-Pro-0813 is the official, non-preview release of DeepSeek-V4-Pro, an open-weight model from DeepSeek AI built for coding and agentic tasks. It replaces the earlier DeepSeek-V4-Pro Preview and adds a speculative decoding module called DSpark for faster inference. According to its model card, it beats the preview version across every listed benchmark and lands in the same performance tier as closed, proprietary models like Anthropic’s Opus-4.8 and other frontier systems referenced in the card as GLM-5.2, Kimi K3, and Fable-5. TL;DR DeepSeek-V4-Pro-0813 is a full release not a preview that supersedes DeepSeek-V4-Pro Preview , with the model card reporting gains on every benchmark it lists.- On Terminal Bench 2.1 , it scores 87.9, just behind Kimi K3 88.3 and Fable-5 88.0 , and ahead of Opus-4.8 85.0 and GLM-5.2 81.0 . - On Humanity’s Last Exam HLE , it scores 42.7 without tools and 60.0 with tools, a large jump that puts its tool-assisted score above Opus-4.8’s 57.9. - The model ships with DSpark speculative decoding , a single-flag feature in both vLLM and SGLang that speeds up generation without needing a separate draft model. - On coding-agent tests like DeepSWE and AutomationBench , the jump from Preview to 0813 is dramatic, DeepSWE goes from 12.8 to 62.7, suggesting the earlier preview was weak specifically at long-horizon coding-agent work. - The model card does not include comparisons against every possible competitor for every benchmark; some cells like Kimi K3 on NL2Repo or Fable-5 on Agents’ Last Exam are simply left blank, meaning no head-to-head number was published. - It runs with an MIT license and recommends a maximum output length of 384K tokens at high reasoning effort levels. One coffee. One working app. You bring the idea. Remy manages the project. How does DeepSeek-V4-Pro-0813 compare to Opus-4.8? The two trade wins depending on the task category. Opus-4.8 leads clearly on raw knowledge and reasoning: it scores 49.8 without tools and 57.9 with tools on HLE, versus DeepSeek’s 42.7 and 60.0. Notice the crossover: DeepSeek actually edges ahead once tools are allowed, which suggests its tool-use integration compensates for a knowledge gap in its base reasoning. Opus-4.8 also wins on NL2Repo 69.7 vs 61.5 and DSBench-Hard 71.7 vs 67.2 , both of which test difficult, structured coding problems. But DeepSeek-V4-Pro-0813 pulls ahead on Terminal Bench 2.1 87.9 vs 85.0 , Cybergym 83.3 vs 78.3 , and AutomationBench Public 31.8 vs 27.2 . The pattern that emerges from the model card: Opus-4.8 tends to hold an edge on tasks requiring broad world knowledge or complex repository-level coding, while DeepSeek-V4-Pro-0813 is more competitive, and sometimes ahead, on terminal-based and automation-heavy agentic work. How does it compare to Kimi K3 and GLM-5.2? Kimi K3 is the strongest competitor across the table for DeepSeek-V4-Pro-0813. It beats DeepSeek on Terminal Bench 2.1 88.3 vs 87.9 , HLE with tools 56.0 vs 60.0, actually DeepSeek wins this one , DeepSWE 67.5 vs 62.7 , Toolathlon-Verified 76.5 vs 74.1 , Agents’ Last Exam 27.6 vs 25.7 , and both DSBench variants. DeepSeek does hold an advantage on Cybergym 83.3 vs 80.0 and AutomationBench Public 31.8 vs 30.8 . Overall, Kimi K3 looks like the closest rival to DeepSeek-V4-Pro-0813 among the models listed, with most gaps sitting within a few points either way. GLM-5.2 trails both models on nearly every benchmark where it appears. Its Terminal Bench 2.1 score 81.0 is notably behind DeepSeek’s 87.9, and its DeepSWE score 46.2 sits well below DeepSeek’s 62.7. GLM-5.2 has no listed score for Cybergym in the model card. Fable-5 with fallback is the wildcard in the table. It posts the highest HLE scores of any model listed 53.3 without tools, 63.0 with tools and leads on several coding benchmarks including DeepSWE 70.0 and DSBench-FullStack 77.2 . But it has no reported score on NL2Repo, Agents’ Last Exam, or AutomationBench, so a full picture of its standing isn’t available from this table. What is DSpark speculative decoding? DSpark is the speculative decoding method attached to DeepSeek-V4-Pro-0813 to speed up inference. Speculative decoding generally works by having a smaller, faster “draft” model propose several tokens ahead, which the full model then verifies in a single pass rather than generating token by token. What’s notable about DSpark, per the model card, is that it doesn’t require a separate draft model checkpoint. The target and draft weights come from the same model file. In practice, enabling it is a single configuration flag. On vLLM, it’s a --speculative-config flag specifying "method":"dspark" along with a token count for how many speculative tokens to generate per step. On SGLang, it’s --speculative-algorithm DSPARK , again without pointing to a separate draft model path. This is a meaningful operational simplification: teams don’t need to maintain, version, and sync two separate checkpoints just to get speculative decoding speedups. What do the benchmark categories actually measure? Plans first. Then code. Remy writes the spec, manages the build, and ships the app. The model card’s benchmark table spans a mix of knowledge, coding, and agentic evaluation types, which is worth understanding before comparing raw scores: HLE Humanity’s Last Exam tests broad reasoning and knowledge, scored both with and without external tool access. The gap between the two scores shows how much a model relies on tools versus internal knowledge. Terminal Bench 2.1 evaluates a model’s ability to complete tasks inside a terminal environment, a proxy for real developer workflows. NL2Repo measures translating natural-language requirements into working repository-level code changes. Cybergym is a security/cyber-focused benchmark, testing tasks like vulnerability analysis or exploitation in controlled environments. DeepSWE evaluates software-engineering agent performance, likely on realistic bug-fixing or feature-implementation tasks. Toolathlon-Verified and AutomationBench test multi-step tool use and automation pipelines. Agents’ Last Exam appears to be an agentic-reasoning analog to HLE, testing long-horizon agent decision making. DSBench-FullStack and DSBench-Hard are internal DeepSeek test sets for full-stack and difficult coding-agent problems, meaning they aren’t independently reproducible outside DeepSeek’s own evaluation harness. The model card notes that for the code-agent tasks, DeepSeek-V4-Pro-0813 was evaluated using the “minimal mode” of DeepSeek Harness their own agent framework at max reasoning effort with temperature = 1.0, top p = 0.95 . That matters for interpretation: harness choice, reasoning effort level, and sampling parameters can all shift scores, so comparisons against other labs’ self-reported numbers should be read with that caveat in mind. Is DeepSeek-V4-Pro-0813 worth using over the preview version? Based on the reported numbers, yes, the gap between DeepSeek-V4-Pro-0813 and DeepSeek-V4-Pro Preview is large enough to make the update a clear improvement rather than an incremental patch. Terminal Bench 2.1 jumps from 72.1 to 87.9. DeepSWE goes from 12.8 to 62.7, a more than fourfold increase. Cybergym rises from 52.7 to 83.3. These aren’t small deltas, they suggest the preview version was materially weaker at sustained, tool-using agentic work, and that the 0813 release specifically targeted that weakness. The model card frames this directly: it says the improvements are “especially pronounced in production environments,” language that lines up with the biggest score jumps showing up in agentic and coding-agent benchmarks rather than raw knowledge tests like HLE where the improvement, from 37.7 to 42.7 without tools, is real but far more modest . For teams already running DeepSeek-V4-Pro Preview in production coding-agent pipelines, the benchmark deltas suggest a meaningful reason to upgrade. For teams evaluating from scratch against Opus-4.8 or Kimi K3, the choice depends on the task: DeepSeek-V4-Pro-0813 is competitive on terminal and automation-heavy work but trails on tasks demanding deep knowledge without tool access. Frequently Asked Questions What does the “0813” in the model name mean? It denotes the release date stamp DeepSeek uses for its model checkpoints, distinguishing this specific version from other DeepSeek-V4-Pro releases like the preview or the Flash variant. Does DeepSeek-V4-Pro-0813 require special hardware to run? The model card’s example deployment commands reference a 4-GPU node using GB300 hardware with FP8 KV-cache and expert parallelism, though it links to separate vLLM and SGLang recipes for other hardware configurations. What license is DeepSeek-V4-Pro-0813 released under? It’s released under the MIT License, covering both the repository code and the model weights. How does DeepSeek-V4-Pro-0813 handle reasoning effort settings? It supports three reasoning effort levels, low, high, and max, which control how much internal deliberation the model performs before producing an answer. The model card recommends a maximum output length of 384K tokens when using the high or max settings. Is DeepSeek-V4-Pro-0813 better than Kimi K3? It depends on the task. Kimi K3 scores higher on Terminal Bench 2.1, DeepSWE, and both DSBench variants, while DeepSeek-V4-Pro-0813 scores higher on HLE with tools and Cybergym. The two models are close across most listed benchmarks.