Published benchmark results of DeepSeek V4 Flash 0731 (open weight, 2026)

The benchmark results published by their vendors for two 2026 open weight models: DeepSeek V4 Flash 0731 (284B / 13B active, 1M context, FP4+FP8 weights) and GLM 5.2 (744B). Terminal Bench, NL2Repo, Cybergym, DeepSWE and Toolathlon on the DeepSeek side; GPQA Diamond and Humanity's Last Exam on the GLM 5.2 side. No estimates: 'not published' where the vendor does not publish the measure.

By QDNA · Published on 2 August 2026 · Updated on 11 October 2026 · 8 min read

Updated on 11 October 2026: local costs of DeepSeek V4 Flash 0731 aligned with the DGX Spark reference prices adopted on 8 October 2026: €2.4 to €42/M depending on throughput and load, €9.4/M under the median assumption.

Updated on 8 October 2026: Artificial Analysis index v4.3.2 re-read on 8 October 2026 (GLM 5.3 at 44.8).

Updated on 26 September 2026: DeepSeek price list of 26 September 2026 (V4 Flash 0731 retired from the API on 10 September, successor V4.1 Flash), V4 Pro 0813 revision, pointer to GLM 5.3 which succeeds GLM 5.2; vendor-published scores unchanged.

Updated on September 2, 2026: GLM 5.2 Terminal Bench 2.1 score split by harness (81.0 on Terminus-2, 82.7 with the best reported harness) instead of a tie with DeepSeek; arXiv 2602.15763 is the GLM-5 report, cited by the GLM 5.2 card; deployment recipe corrected (the 167 GB official checkpoint does not fit in a single Spark).

Methodology

This article reports only scores published by the vendors on their model cards and official reports. DeepSeek and Zhipu AI (GLM) are, to our knowledge, the only two vendors publishing these benchmarks for the exact checkpoints covered here. Where a vendor does not publish a measure, the cell reads 'n.p.'

(not published): no value is estimated, copied from a third party, or compared against a closed model whose vendor does not publish the same benchmarks. For a consistent, independent cross-model comparison, we point readers to the Artificial Analysis and LiveBench panels.

The sources are: arXiv 2606.19348 (DeepSeek V4 technical report), DeepSeek V4 Flash 0731 Hugging Face model card, GLM 5.2 Hugging Face model card and GLM-5 technical report (arXiv 2602.15763), cited by the GLM 5.2 model card. Model cards re-read on 2 September 2026.

Published official benchmarks

n.p. = not published by the vendor. Scores from the official model cards / reports: DeepSeek (Hugging Face), GLM 5.2 (Hugging Face, arXiv report 2602.15763). For an independent cross-model comparison: Artificial Analysis and LiveBench.

Reading. DeepSeek publishes its results on agentic and coding benchmarks (Terminal Bench, NL2Repo, Cybergym, DeepSWE, Toolathlon), whereas GLM 5.2 mainly documents knowledge and scientific reasoning (GPQA Diamond, Humanity's Last Exam). On Terminal Bench 2.1, the only benchmark common to both cards, GLM 5.2 publishes 81.0 with the Terminus-2 harness and 82.7 with the best reported harness (Claude Code), while DeepSeek publishes 82.7 with its own harness: three different conditions, which rule out reading an equality into it. An 'n.p.' cell does not signal a weak result but the absence of a measure published by the vendor; filling it with an estimate would mean inventing a figure. To position these models against others, open or closed, on a single methodology, see Artificial Analysis or LiveBench.

Recommendations by use case

Reading. For a sovereign on-site deployment, DeepSeek V4 Flash 0731 is the default choice: open weights under the MIT licence, 1M context and the most complete set of published agentic benchmarks. GLM 5.2 rounds out the picture on knowledge and scientific reasoning.

Hardware sizing (GPU, memory, throughput, concurrent users) is covered in our DeepSeek deployment guide for DGX Station and Spark clusters.

Licence and sovereignty

Both models ship under the MIT licence with open weights: fully on-site execution, no data transfer outside the infrastructure, no dependency on a third-party API and no deactivation risk. That is the basis for GDPR, HDS and SecNumCloud compliance on sensitive data.

Conclusion

On the benchmarks their vendors publish, DeepSeek V4 Flash 0731 documents the agentic and coding side (Terminal Bench 2.1 82.7, NL2Repo 54.2, DeepSWE 54.4, Toolathlon 70.3, Cybergym 76.7), while GLM 5.2 documents the knowledge and reasoning side (GPQA Diamond 91.2, Humanity's Last Exam 40.5).

Both are open weight under the MIT licence, so deployable on-premises. We do not rank them against proprietary models: those do not publish the same measures, and any quantified comparison goes through an independent single-methodology panel such as Artificial Analysis or LiveBench.

References

• DeepSeek V4 Flash 0731, dépôt officiel Hugging Face
• Rapport technique DeepSeek V4 (arXiv 2606.19348)
• Artificial Analysis Intelligence Index v4.1
• LiveBench open weight rankings
• OpenRouter model rankings
• Unsloth, documentation DeepSeek V4
• Fiche QDNA : DeepSeek V4
• Fiche QDNA : vLLM
• Fiche QDNA : DGX Station
• Fiche QDNA : DGX Spark