先看 DeepSeek 自有信号：官网公开 X 入口、官方 API 文档变化，以及会影响 V4 Pro / V4 Flash 接入方式的开发者更新。Reasonix、社区、市场和传闻内容只做次级补充标注。
Checked on July 6, 2026: DeepSeek's official docs currently expose both a Python thinking-mode tool-call sample and a separate output transcript page, giving developers a first-party reference for tool schemas, chained calls, blank content turns, and reasoning_content replay behavior.
Checked on July 5, 2026: DeepSeek's official homepage still anchors the official `@deepseek_ai` X presence, but the strongest current first-party developer signal is in the API docs and changelog, which both keep the July 24, 2026 retirement warning for `deepseek-chat` and `deepseek-reasoner` while pointing users to explicit `deepseek-v4-flash` and `deepseek-v4-pro` model names.
截至 2026 年 7 月 2 日，OpenAI Codex 文档已经写明 custom model providers：用 `model_provider`、`model_providers`、`base_url` 和 `env_key` 配置第三方模型。本站这条新闻已改成更准确的说法：DeepSeek 可以作为 Codex 第三方 provider，用我们的 DeepSeek 接入即可；同时保留 Responses API、profile、Desktop 模型选择和 V4 thinking mode 的验证提醒。
Checked on July 2, 2026: DeepSeek's official API reference documents a `GET /models` endpoint that currently examples only `deepseek-v4-flash` and `deepseek-v4-pro`, plus a `GET /user/balance` endpoint with `is_available`, currency, granted balance, and topped-up balance fields for direct budget checks.
Checked on July 1, 2026: DeepSeek's official X account published a permanent-discount update, and the current official Models & Pricing page still shows the lower DeepSeek V4 Flash and Pro token rates together with the existing 1M-context and 384K max-output contract.
Checked on June 30, 2026: DeepSeek's official Token & Token Usage page says tokens are the billing unit, gives rough heuristics of about 0.3 token per English character and 0.6 token per Chinese character, and links a `deepseek_tokenizer.zip` package for offline token estimation while still treating API `usage` output as the real source of truth.
截至 2026 年 6 月 30 日，澎湃新闻与财联社报道均指向同一调整：DeepSeek V4 正式版计划 7 月中旬上线，并在北京时间 9:00-12:00、14:00-18:00 引入高峰计费；官方公开价格页目前仍显示平时价格表，因此本站将其写成报道类价格更新，不改 /pricing 库存卡片。
Checked on June 29, 2026: DeepSeek's official Tool Calls guide documents OpenAI-compatible function calling, a Beta-only `strict` mode for exact schema matching, concrete supported and unsupported JSON Schema rules, and a required `tool` role replay pattern after execution.
Checked on June 28, 2026: DeepSeek's official Models & Pricing and Rate Limit pages currently show a 1M context window, 384K max output, per-1M token pricing, and account-level concurrency limits of 2500 for DeepSeek V4 Flash and 500 for DeepSeek V4 Pro, with 429s when that cap is exceeded.
Checked on June 27, 2026: DeepSeek's official multi-round conversation guide says `/chat/completions` does not store context server-side, so each follow-up call must resend the prior conversation history, including earlier assistant output, before the next user turn.
Checked on June 26, 2026: DeepSeek's official WorkBuddy or CodeBuddy page now documents exact user-level and project-level `models.json` paths, `https://api.deepseek.com/v1/chat/completions` as the endpoint, a UTF-8 without BOM requirement, and a restart-first troubleshooting flow for V4 Pro and V4 Flash.
Checked on June 25, 2026: DeepSeek's official Anthropic API page says `claude-opus*` maps to V4 Pro, `claude-haiku*` and `claude-sonnet*` map to V4 Flash, unsupported model names automatically fall back to Flash, and the mapping can help bypass Claude Desktop developer-mode model-name restrictions.
Checked on June 24, 2026: DeepSeek's official GitHub Copilot CLI page says teams should use the Anthropic-compatible endpoint, warns that the OpenAI provider triggers a `400` reasoning-content error, sets explicit token ceilings for `deepseek-v4-pro`, and clarifies that offline mode still sends prompts to DeepSeek.
Checked on June 23, 2026: DeepSeek's official Claude Code page now documents the Anthropic-compatible environment contract, native Web Search support, and an explicit mapping where Claude-style model names route to DeepSeek V4 Pro or Flash.
Checked on June 22, 2026: DeepSeek's official OpenClaw page now documents `openclaw onboard --install-daemon`, QuickStart provider selection, manual entry of `deepseek-v4-pro` or `deepseek-v4-flash`, and follow-up launch routes through the dashboard, TUI, or terminal.
Checked on June 21, 2026: DeepSeek's official AstrBot page now documents uv and Docker installs, a local web UI at `localhost:6185`, and a direct provider flow where users add DeepSeek, paste a DeepSeek API key, and set the default chat model inside AstrBot.
Checked on June 20, 2026: DeepSeek's official Reasonix integration page now documents `npx reasonix code`, stores the API key in `~/.reasonix/config.json`, defaults to DeepSeek V4 Flash, and exposes `/pro` plus `/preset max` for stronger reasoning when needed.
Checked on June 19, 2026: DeepSeek's official Oh My Pi integration page says the built-in DeepSeek entries are not reliable enough on their own, requires a custom `~/.omp/agent/models.yml`, and documents the exact compat fields needed to avoid 400 errors in thinking-mode tool-call sessions.
Checked on June 18, 2026: DeepSeek's official Pi integration guide now gives a concrete `models.json` provider block with `api.deepseek.com`, 1M context metadata, Pro and Flash pricing fields, and a reasoning-effort map that translates Pi's levels to DeepSeek's supported behavior.
截至 2026 年 6 月 18 日，DeepSeek 官方研究与 GitHub 页面已经能确认 Janus-Pro、DeepSeek-VL2 等多模态能力；但官方 V4 API 文档仍把 Pro 和 Flash 写成文本主路由，Copilot 文档也说明 V4 本身是 text-only，因此应写成多模态生态扩展，而不是已确认原生 V4 Vision API。
Checked on June 17, 2026: DeepSeek's official OpenCode integration guide tells existing users to upgrade to at least v1.14.24, switch providers through `/connect`, enter a DeepSeek API key, and select DeepSeek V4 Pro instead of relying on custom endpoint hacks.
Checked on June 17, 2026: WSJ, The Information, Wallstreetcn, 36Kr, TMTPost, and other outlets now report that DeepSeek has closed a first external funding round of more than 50 billion yuan, with Liang Wenfeng investing 20 billion yuan, Tencent 10 billion yuan, and strict control terms around a Liang-managed limited partnership.
Checked on June 16, 2026: a Linux.do thread and follow-up Chinese tech coverage say some users may be seeing a DeepSeek web Expert Mode gray test with fresher knowledge-cutoff answers and much stronger SVG / Three.js-style coding output. Official DeepSeek docs still list only V4 Preview, V4 Pro, and V4 Flash, so this should be treated as a community signal rather than an official V4.1 release.
Checked on June 16, 2026: DeepSeek's official GitHub Copilot CLI guide now tells teams to route BYOK through the Anthropic-compatible endpoint, explains why the OpenAI provider path triggers a `400` reasoning-content error, and documents explicit prompt and output token caps for DeepSeek V4 Pro.
Checked on June 15, 2026: DeepSeek's official Hermes Agent integration page now documents a curl install flow, Quick Setup with DeepSeek as the provider, `https://api.deepseek.com` as the base URL, and `deepseek-v4-pro` as the default model choice.
Checked on June 14, 2026: DeepSeek's official GitHub Copilot integration page now documents a VS Code extension that adds DeepSeek V4 Pro and Flash to Copilot Chat, keeps agent mode, tool calling, skills, and MCP, and adds an optional vision proxy path.
截至 2026 年 6 月 11 日，DeepSeek 官方文档已给出更清晰的生产运行基线：V4 Pro 账户级并发 500、V4 Flash 2500，并明确说明 `user_id` 隔离和请求 keep-alive 的处理方式。
截至 2026 年 6 月 10 日，DeepSeek 官方 Claude Code 页面已直接说明其 API 原生支持 Claude Code 的 Web Search；同页仍把主模型固定为 V4 Pro，把子代理固定为 V4 Flash。
截至 2026 年 6 月 9 日，DeepSeek 英文首页仍直接链接到 @deepseek_ai 的 V4 Preview 帖子；与此同时，官方 API 文档继续把 V4 Pro / V4 Flash 作为当前模型名，并保留 7 月 24 日旧别名退役提醒。
今天可发布的是本地部署主线更新：`teamblobfish/DeepSeek-V4-Flash-GGUF` 现在把命名好的 GGUF 文件、`llama.cpp`/Ollama/vLLM 命令、Apple Silicon 吞吐数据，以及必须使用的 `cchuter/llama.cpp` V4 分支绑在一起，证据强于旧的泛化量化列表。
今天可发布的只有 TUI 主线更新：公开维护中的 CodeWhale 页面已把 v0.8.50 作为稳定基线，继续把 `codewhale` 设为规范命令，并强调 `.codewhale` 与校验过的 release 制品才是当前推荐安装路径。
今天可发布的是一条双信号开发者更新：DeepSeek TUI 已前进到 CodeWhale v0.8.49，而 DeepSeek 官方 Anthropic API 文档现在明确写出了 Claude 风格模型名到 DeepSeek Pro / Flash 的映射关系。
今天可发布的开发者信号仍来自 TUI 主线：最新稳定版已到 v0.8.47，而且 README 与 Docker 持久化示例把主配置目录改为 `.codewhale`，`.deepseek` 只保留兼容回退角色。
今天可发布的开发者信号仍然来自 TUI 主线：GitHub Releases 已更新到 CodeWhale v0.8.43，而本地部署与 Claude Code 两条线今天没有足够新的官方变化，不应硬写新结论。
Today's accepted developer signal stays on the TUI track: GitHub Releases now marks CodeWhale v0.8.42 as latest, makes `codewhale` canonical, and limits legacy `deepseek` compatibility binaries to a temporary migration bridge.
DeepSeek 官方价格页已说明，V4-Pro 的 75% 折扣窗口结束后，API 价格会正式调整为原定价的 1/4。中文页面应写成官方价格重置，而不是承诺永远不变。
Today's accepted developer signal stays on the TUI track: the public DeepSeek TUI website still points users at `deepseek`, but the live activity feed now shows CodeWhale rename work across docs, npm, and binaries while keeping compatibility aliases.
Today's accepted developer signal stays on the TUI track: the official release line has moved to v0.8.40, and the release/install surface now gives one current stable reference across npm, GHCR, Cargo, manual binaries, and published checksums.
Today's accepted developer signal stays on the TUI track: the official README now ties install-specific upgrade commands, a persistent Docker home volume, beta-default DeepSeek routing, and SHA-256 download checks into one clearer operational story.
Today's accepted developer signal stays on the TUI track: the official README and install docs now clarify `deepseek auth status`, credential precedence, `auth clear`, a documented Nix path, and how auto routing is billed against the model that actually ran.
两篇中文报道都指向同一信号：DeepSeek 正在招聘/组建 Agent Harness 方向，把模型能力接入桌面端、终端、文件读写和工具调用等真实开发工作流。但这仍是报道和招聘信号，不是官方产品发布。
Today's accepted developer signal is on the Claude Code track: DeepSeek's official docs now tie Claude-style integrations more clearly to exact V4 model strings, while the main API quick-start warns that `deepseek-chat` and `deepseek-reasoner` will be deprecated on July 24, 2026.
Today's accepted developer signal stays on the TUI track: the official release line has moved to v0.8.39, and the install docs now make mirror variables, forked release sources, CI-safe optional installs, Scoop lag, and broader source-build paths much easier to document accurately.
Today's accepted developer signal is again narrowly about DeepSeek TUI: official install docs now document Homebrew, raise Cargo/source-build expectations to Rust 1.88, show checksum verification and China-friendly mirror variables, and expose a local runtime contract built around `deepseek doctor --json` plus `deepseek serve --http`.
Today's publishable developer update is narrowly scoped to DeepSeek TUI: the project now presents v0.8.28 as latest stable, documents `deepseek doctor`, adds Docker as an official install path, expands skill discovery details, and tells users that only the latest stable release receives security patches.
Today's accepted update is a compact developer round-up: DeepSeek TUI v0.8.17 tightened paired-binary install guidance and corrected the `deepseek-cn` host, DeepSeek's Claude Code docs now explicitly pin the subagent model, and the official V4 Flash Hugging Face card exposes copyable vLLM and SGLang local server commands.
这是一条报道类消息，不是官方发布。重点是跟踪 V4.1 的 6 月窗口、视觉/音频输入、MCP 支持和融资估值区间。
部分用户已看到 DeepSeek 视觉能力灰度入口。技术报告强调点和框等视觉原语，让模型在推理过程中绑定图像位置。
缓存命中输入价格极低，适合固定 system prompt、长上下文复用和高频开发者工作流，是 DeepSeek V4 成本优势的关键。
DeepSeek V4 Pro 降价是当前价格主线。本站 Coding Plan 是一次性库存产品，token 官方价格与购买卡片价格需要分开阅读。
本地 Flash 是重要部署信号，但生产场景仍要区分社区验证、本地实验和官方云端 API。
本地部署适合实验和隐私敏感场景；稳定生产、并发和成本核算仍需要官方 API 作为基线。
关注 DeepSeek V4 在 OpenClaw PinchBench、编程、推理和工具调用中的表现，以及它与 GPT、Claude、Gemini 的差距。
市场信号应与官方发布分开阅读。真正影响用户决策的是价格、可用性、API 稳定性和工具链兼容。
中文社区对 DeepSeek V4 的关注集中在降价、API 接入、Claude/开发工具映射和本地部署可行性。
整理 DeepSeek V4 的关键规格信号，并提醒读者以官方文档和实际 API 行为为准。
汇总 DeepSeek V4 的发布定位、基准测试、迁移指南和本站有库存 Coding Plan 的购买边界。
A new commit titled "act_quant_kernel" merged into the DeepSeek-V3 repository, indicating ongoing low-level kernel optimizations for RMSNorm operations.
DeepSeek's GitHub repository now supports the 'scale_fmt=ue8m0' quantization format, indicating potential low-bit inference optimizations.
DeepSeek's official GitHub repository has committed fixes for RMSNorm and activation quantization kernels, indicating ongoing low-level optimization work.
A commit fixes the `act_quant_kernel` in the DeepSeek-V3 codebase, indicating ongoing low-level optimization work on the model's quantization implementation.
DeepSeek has officially archived the DeepSeek-V3 model with a v1.0.0 release, primarily for generating a permanent DOI.