I now have the arxiv technical report abstract, which is the primary technical source. I have all the key information I need. Let me compile my findings. I have gathered comprehensive information from: 1. **ARC-AGI results page** (primary benchmark source) - verified scores 2. **arXiv technical report** (2606.19348) - architecture details 3. **Artificial Analysis article** - independent benchmark analysis 4. **Shadowfetch news** - release details Let me now write the final answer. --- # DeepSeek V4 Flash 0731 — Investigation Summary ## What it is **DeepSeek V4 Flash 0731** is a checkpoint released by DeepSeek on **July 31, 2026**, described as the *official* release of the DeepSeek-V4-Flash model, superseding the earlier preview. It is the "Flash" (smaller, cheaper) member of the DeepSeek-V4 series, paired with the larger DeepSeek-V4-Pro (1.6T params). The Flash model has **284B total parameters with 13B active** at inference (Mixture-of-Experts), a **1M-token context window**, and text-only input/output. It ships in **three reasoning-effort variants: Max, High, and Low** ([arXiv:2606.19348](https://arxiv.org/abs/2606.19348); [Shadowfetch](https://www.shadowfetch.com/news-room/deepseek-v4-flash-0731-agentic-release)). ## What changed vs. the previous DeepSeek V4 Flash (April 2026) - **Architecture is unchanged** — same 284B/13B MoE structure, same 1M context, same pricing ($0.14/$0.28 per 1M input/output tokens). The 0731 checkpoint is a *post-training / capability* upgrade, not a new architecture ([Artificial Analysis](https://artificialanalysis.ai/articles/deepseek-v4-flash-0731-scores-50-on-the-artificial-analysis-intelligence-index-10-points-above-previous-deepseek-v4-flash)). - **Intelligence Index: 50, up 10 points** from the previous Flash (40), putting it 6 points ahead of DeepSeek V4 Pro and within 1 point of GPT-5.6 Luna (max, 51) and GLM-5.2 (max, 51). It remains 7 points behind the open-weights frontier, Kimi K3 (max, 57) ([Artificial Analysis](https://artificialanalysis.ai/articles/deepseek-v4-flash-0731-scores-50-on-the-artificial-analysis-intelligence-index-10-points-above-previous-deepseek-v4-flash)). - **Agentic performance jumped**: GDPval-AA v2 Elo rose from **1189 → 1559**; Terminal-Bench 2.1 rose 17 points to 79%; τ³-Bench Banking rose 8 points to 31% ([Artificial Analysis](https://artificialanalysis.ai/articles/deepseek-v4-flash-0731-scores-50-on-the-artificial-analysis-intelligence-index-10-points-above-previous-deepseek-v4-flash)). - **Fewer hallucinations**: AA-Omniscience Index improved from −23 to −16, driven *entirely* by a lower hallucination rate (84%, down 12 points) — accuracy was unchanged at 37% ([Artificial Analysis](https://artificialanalysis.ai/articles/deepseek-v4-flash-0731-scores-50-on-the-artificial-analysis-intelligence-index-10-points-above-previous-deepseek-v4-flash)). - **Improved on every Intelligence Index eval**: CritPt +9 (to 17%), SciCode +5 (to 50%), Humanity's Last Exam +5 (to 37%), AA-LCR +3 (to 66%), GPQA Diamond +1 (to 91%). It also used ~12% fewer output tokens (~206M vs ~234M) ([Artificial Analysis](https://artificialanalysis.ai/articles/deepseek-v4-flash-0731-scores-50-on-the-artificial-analysis-intelligence-index-10-points-above-previous-deepseek-v4-flash)). - **Release/format changes**: OpenAI-compatible reasoning controls (low/high/max effort), a dedicated encoding workflow, DSpark-enabled vLLM/SGLang serving guidance, and a recommended 384K max output length for high/max reasoning in agentic workflows ([Shadowfetch](https://www.shadowfetch.com/news-room/deepseek-v4-flash-0731-agentic-release)). ## ARC-AGI results (verified by ARC Prize, July 31, 2026) From the [ARC-AGI results page](https://arcprize.org/results/deepseek-v4-flash-0731), the verified Semi-Private scores are: | Variant | ARC-AGI-1 | ARC-AGI-2 | ARC-AGI-3 | |---------|-----------|-----------|-----------| | **Max** | **89.0%** | **61.4%** | — | | High | 87.0% | 56.0% | — | | Low | 84.0% | 46.0% | — | At max effort it scores **89.0% on ARC-AGI-1 Semi-Private at $0.02/task** and **61.4% on ARC-AGI-2 Semi-Private at $0.04/task**. The page also publishes per-task pass/fail for all 120 ARC-AGI-2 and 400 ARC-AGI-1 public-eval tasks across the three reasoning levels. ## Why it matters 1. **State-of-the-art open-weights reasoning at low cost.** The 0731 checkpoint is the strongest open-weights performer on ARC-AGI-2 (61.4% Semi-Private) at just $0.04/task, and it lands on Artificial Analysis' Pareto frontier for Intelligence vs. Cost-per-Task ([ARC-AGI](https://arcprize.org/results/deepseek-v4-flash-0731); [Artificial Analysis](https://artificialanalysis.ai/articles/deepseek-v4-flash-0731-scores-50-on-the-artificial-analysis-intelligence-index-10-points-above-previous-deepseek-v4-flash)). 2. **Cost advantage via aggressive cache pricing.** Its first-party API cost-per-task is ~60% lower than the comparable-intelligence GPT-5.6 Luna (max), driven by a ~98% cache-hit discount (vs. the industry-standard ~90%) ([Artificial Analysis](https://artificialanalysis.ai/articles/deepseek-v4-flash-0731-scores-50-on-the-artificial-analysis-intelligence-index-10-points-above-previous-deepseek-v4-flash)). 3. **Agentic reasoning is the headline improvement.** The 0731 build is explicitly positioned as an agentic-capability release, with the largest gains in agentic benchmarks (GDPval-AA +370 Elo) and reduced hallucination — the key failure mode for agentic work ([Artificial Analysis](https://artificialanalysis.ai/articles/deepseek-v4-flash-0731-scores-50-on-the-artificial-analysis-intelligence-index-10-points-above-previous-deepseek-v4-flash); [Shadowfetch](https://www.shadowfetch.com/news-room/deepseek-v4-flash-0731-agentic-release)). 4. **Long-context efficiency is the underlying technical story.** The V4 series introduces Compressed Sparse Attention (CSA) + Heavily Compressed Attention (HCA), Manifold-Constrained Hyper-Connections (mHC), and the Muon optimizer, pre-trained on 32T+ tokens. In 1M-token contexts, V4-Pro needs only 27% of single-token inference FLOPs and 10% of the KV cache vs. DeepSeek-V3.2 — making million-token, long-horizon agentic tasks practical ([arXiv:2606.19348](https://arxiv.org/abs/2606.19348)). ## Caveats / what's missing - **Weights were not yet public at the time of these reports** — Artificial Analysis notes DeepSeek is "expected to release the model's full weights in the coming weeks," so open-weights claims are prospective ([Artificial Analysis](https://artificialanalysis.ai/articles/deepseek-v4-flash-0731-scores-50-on-the-artificial-analysis-intelligence-index-10-points-above-previous-deepseek-v4-flash)). - The **arXiv paper (2606.19348) is the V4-series technical report** ("Towards Highly Efficient Million-Token Context Intelligence," submitted Apr 26, 2026) — it describes the architecture but is the *preview* report; the 0731 checkpoint-specific details come from the model card and third-party analyses. - I could not render the HuggingFace model card directly in this session, so release-card specifics (beyond what Shadowfetch and Artificial Analysis report) are drawn from those secondary sources. **Strongest sources:** [ARC-AGI results page](https://arcprize.org/results/deepseek-v4-flash-0731) (verified benchmark scores), [arXiv:2606.19348](https://arxiv.org/abs/2606.19348) (architecture/technical report), [Artificial Analysis article](https://artificialanalysis.ai/articles/deepseek-v4-flash-0731-scores-50-on-the-artificial-analysis-intelligence-index-10-points-above-previous-deepseek-v4-flash) (independent benchmark comparison), and [Shadowfetch release coverage](https://www.shadowfetch.com/news-room/deepseek-v4-flash-0731-agentic-release) (release/format details).