DeepSeek V4 Flash 0731

284B total, 13B active per token (MoE). Same FP4+FP8 hybrid-attention family as V4 Pro: Compressed Sparse Attention (CSA) + Heavily Compressed Attention (HCA) across 61 layers, manifold-constrained Hyper-Connections (mHC), Muon optimizer, 32T+ pretraining tokens.

• Context: 1M native, 384K max output; three modes (non-think / think-high / think-max).

The 2026-07-31 iterative update of V4 Flash (supersedes the April model). Per Artificial Analysis, a 10-point Intelligence Index jump to 50 - 6 points above V4 Pro, 1 behind GLM 5.2 / GPT-5.6 Luna, 7 behind Kimi K3. Agentic Elo 1559 (up from 1189), Terminal-Bench 2.1 79% (+17), Humanity’s Last Exam 37% (+5), GPQA-Diamond 91% (+1), SciCode 50% (+5). Token usage -12%; hallucination rate 84% (a 12-point drop); AA-Omniscience Index -16 (+7). Pricing unchanged at $0.14/$0.28 per 1M in/out (cache-hit $0.0028/M, a 98% discount). The model the deepseek-chat/reasoner API aliases now route to (retired 2026-07-24).

• Local run (2026-07-31): Unsloth’s Dynamic 2.0 GGUFs landed local inference. UD-Q4_K_XL is a 155GB lossless 4-bit build (~168GB RAM); UD-Q8_K_XL is a 162GB 8-bit full-precision build (~175GB RAM, only 7GB bigger than Q4 because the 13B active experts dominate). Both fit two stacked DGX Sparks (256GB unified via ConnectX-7) or a 192GB+ unified rig; a smaller 3-bit (~110GB RAM) that would fit a single 128GB Spark is announced but not yet published. Run via Unsloth or llama.cpp -hf - Ollama only ships the :cloud endpoint, so there is no local Ollama tag.

Open weights under MIT (full weights expected in the coming weeks per DeepSeek; the Unsloth GGUFs are available now).

Agents on Rails benchmark (Aug 2026, Le Mans round). 65.1% accuracy on 63 runs - last of the 16-model field, with only 12.7% API recall. The low recall suggests it often hand-rolled fixes instead of using the relevant Rails API. Still the cheapest per-token model in the field, so the gap is cost-vs-accuracy rather than capability at any price.