OpenCode’s Go documentation lists DeepSeek V4 Flash in the provider’s current model roster, alongside DeepSeek V4 Pro and DeepSeek V4 Flash Vision Exp. The same page describes Go as an optional $10-per-month subscription for low-cost access to open coding models, with service intended primarily for international users To use Go, developers sign in to OpenCode Zen, subscribe, copy an API key, run /connect in the OpenCode TUI, select OpenCode Go, and then run /models to view the available models. The documentation says Go also works with other coding agents that make similar requests, but warns that the model lineup can change as Open

DeepSeek released a public beta of its V4-Flash API on July 31, 2026 under the build designation V4-Flash-0731, and the change is a re-run of post-training rather than a new architecture. The model keeps the same 284B-total / 13B-active MoE structure, a one-million-token context window, and the MIT license that permits DeepSeek reports that V4-Flash-0731 now beats its own V4-Pro-Preview across all nine agent and coding benchmarks the company published, including 82.7 on Terminal Bench 2.1 (vs. 72.1 for the Pro-Preview and 61.8 for the prior Flash Preview) and 54.4 on DeepSWE, up from 7.3 on the preview, a roughly 645% jump on a bench

Artificial Analysis reports that the DeepSeek V4 Flash 0731 update (July 31, 2026) scores 50 on the Artificial Analysis Intelligence Index, a 10-point jump over the original April 2026 DeepSeek V4 Flash and 6 points ahead of DeepSeek V4 Pro. It lands 1 point behind GPT-5.6 Luna (max, 51) but comes in at 60% lower Cost Agentic performance is the biggest gainer: GDPval-AA v2 Elo jumps to 1559 from 1189, Terminal-Bench 2.1 rises 17 points to 79%, and τ³-Bench Banking climbs 8 points to 31%. AA-Omniscience improves by +7 to an index of -16, driven entirely by a 12-point drop in hallucination rate (to 84%) with raw accuracy unchanged. De

DeepSeek V4-Flash at Scale: A Benchmark-Driven Deployment Guide A hands-on tutorial comparing Token API, PTU, Model Unit, and Bare Metal GPU for production LLM inference. Real numbers. Real …

The OpenCode changelog records a Core-level fix that explicitly targets the subject model: "Applied the correct sampling defaults to DeepSeek V4 Flash on supported providers." This is first-party provider documentation and is directly relevant to OpenCode Go users, whose provider entry sits among the "supported provide For developers calling deepseek-v4-flash via OpenCode Go's /v1 API, the sampling-defaults correction matters because it governs temperature, top-p, top-k and related parameters that affect determinism, tool-calling behavior, and output quality. Other bundled Core fixes in the same entry touch OpenCode Go paths as well,

DeepSeek's official API changelog documents the 2026-04-24 launch of DeepSeek-V4-Pro and DeepSeek-V4-Flash, served through both the OpenAI ChatCompletions-compatible interface and an Anthropic-compatible interface using the model parameters deepseek-v4-pro and deepseek-v4-flash. The same entry announces that the legacy This changelog entry is significant for developers integrating DeepSeek APIs because the 2026-07-24 sunset date for deepseek-chat and deepseek-reasoner forces a migration to the explicit deepseek-v4-flash or deepseek-v4-pro model names. The documentation also records earlier V3.2 and V3.1 updates, providing historical