V4 Flash-0731
DeepSeek V4 Flash 0731 is a sparse mixture-of-experts model (284B total / 13B active parameters) with a 1M-token context, released as a re-post-trained revision of the V4 Flash preview tuned for coding, reasoning, and agentic workflows. It delivers near-flagship agentic/coding performance at exceptionally low cost.
Empirical Evaluation Results
Architectural Profile & Capabilities
DeepSeek V4 Flash 0731 (OpenRouter slug deepseek/deepseek-v4-flash-0731) is a text-only sparse mixture-of-experts language model from DeepSeek. It carries 284B total parameters with roughly 13B active parameters per token, paired with a 1,048,576-token context window and an attached speculative-decoding module for accelerated inference. The 0731 build is explicitly a re-post-trained revision of the earlier V4 Flash preview: core architecture and size were retained while post-training was reworked to substantially enhance agentic capabilities — coding, tool use, and long-horizon workflows. Notably, it outperforms the larger DeepSeek-V4-Pro (Preview) on several agentic benchmarks despite its far smaller activated parameter count, making it one of the strongest price-performance open models of the current cycle. Pricing sits at $0.008/1M input and $1.28/1M output (OpenRouter), with provider-side floors as low as ~$0.06–$0.22/1M input and ~$0.18–$0.66/1M output depending on cache-hit/miss status and host. Its benchmark profile is agentic-heavy: strong results on Terminal-Bench 2.1 (78.7%), Toolathlon Verified (70.7%), AutomationBench (54.0%), TAU-2 Airline (72.8%), and GPQA Diamond (90.8%), while its harder-frontier scores (Terminal-Bench 4.0 at 12.1%, GDP PDF at 11.0%, Critpt at 16.6%) and very low omniscience/hallucination metrics betray a model optimized for procedural execution rather than broad factual recall. It is best treated as a fast, cheap agentic workhorse rather than a frontier reasoning or long-form knowledge engine. NOTE ON GOVERNANCE: All scores supplied in the Ground-Truth Baseline were preserved verbatim and unmodified. Additional entries below were only appended where a source explicitly attributed them to the exact 0731 variant; no sibling (V4 Pro) or preview-version numbers were used as proxies.
284B total parameters with ~13B active parameters per token (approx. 4.6% activation ratio). 1,048,576-token (1M) input context, text-only modality. Released with an attached speculative-decoding module for latency reduction. The 0731 revision retains the V4 Flash preview's architecture and size and is primarily a re-post-training update targeting agentic capability. The specific attention mechanism (MLA vs. GQA) is NOT publicly confirmed for this exact variant via official model cards/technical reports and should not be inferred from V4 Pro documentation; expert count, routing top-k, hidden size, and layer placement are likewise not disclosed in available exact-variant sources. Test-time reasoning tokens are supported (reasoning/thinking budget configurable).
Recommended Workloads & Primary Use Cases
- •Attention/architecture internals are undisclosed for this exact revision — do not assume MLA or GQA; capacity planning must rely on measured serving benchmarks rather than V4 Pro documentation.
- •Very low long-context retrieval reliability: Mlcr 13.3% and AA-Omniscience Index −14.3 (Non-Hallucination Rate only 8.3%) indicate significant degradation and hallucination risk on recall-heavy or needle-in-haystack tasks near the context limit despite the 1M window.
- •Frontier-hard reasoning is a weakness: Terminal-Bench 4.0 (12.1%), GDP PDF (11.0%), and Critpt (16.6%) show sharply lower performance on the hardest, most novel-task suites versus its strong routine-agentic scores.
- •Substantial cross-provider variance: same weights are served by 24+ providers with differing quantization and hardware, producing materially different throughput, TTFT, and sometimes score differences (e.g., some published scores are labeled 'Max' and may not reflect default endpoint inference settings).
- •Output/reasoning tokens are billed at a much higher rate than input ($1.28/1M output vs $0.008/1M input); extended thinking on agentic loops can rapidly escalate cost, and there is no confirmed explicit cap on reasoning-token expenditure per request.
- •Some leaderboard numbers (Vals AI, DeepSearchQA, BrowseComp) are context- and budget-dependent (turn limits, tool settings) and should not be compared directly to static scientific benchmarks.
Calls to V4 Flash-0731 are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by DeepSeek.
More Models from DeepSeek
Compare other engines available in this laboratory.
Terminal-Bench 2.1: 87.9%, DeepSWE: 62.7%, CyberGym: 83.3%, AutomationBench: 31.8%, Toolathlon-Verified: 74.1%, HLE w/ tools: 60.0%, GPQA Diamond: 90.1%, Artificial Analysis Intelligence Index: 53.2, AA Coding Index: 68.8, AA Agentic Index: 49.6 | The official GA release of V4 Pro, superseding the April preview (released Aug 13, 2026).
V4 Pro is a high-performance foundation model engineered by DeepSeek, accessible with Zero Data Retention on ARMES.
DeepSeek V4.1 Flash 1017 is the October 17, 2026 snapshot of DeepSeek V4.1 Flash — a 552B-parameter sparse Mixture-of-Experts model built on DeepSeek's Causal Encoder-Decoder (CED) architecture, activating only ~8B parameters during prefill and ~16B during decode. It pairs a 1,048,576-token context window with native image input, controllable reasoning effort, MIT-licensed open weights, and an extremely low price ($0.15/1M in, $0.60/1M out), positioning it as a cost-dominant agentic/coding workhorse.