Source-backed DeepSeek changes, sorted by what they affect: API breaking changes first, then model and pricing changes, then workflow and tooling updates. Official @deepseek_ai and DeepSeek docs are the primary sources; Reasonix, local deployment, and Claude Code items are labeled as developer signals, not broad AI market chatter.
Peak hours are 09:00–12:00 and 14:00–18:00 Beijing time (01:00–04:00 and 06:00–10:00 UTC), Monday–Friday, excluding Chinese public holidays. All other hours, weekends, and Chinese public holidays are off-peak. Official API rates are separate from Coding Plan purchases. Verified 2026-10-11. Official prices, including cache-hit rates
Endpoints, rate limits, request schemas, and alias retirements that can break existing calls.
Checked on July 14, 2026: DeepSeek's official docs sitemap currently lists dedicated `chat_curl`, `chat_nodejs`, and `chat_python` sample pages, and all three use the current `https://api.deepseek.com` endpoint plus `deepseek-v4-pro` instead of older alias-first snippets.
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.
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 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.
Checked on June 11, 2026: DeepSeek's official docs now give a cleaner production picture for V4 traffic management, with 500 Pro concurrency, 2500 Flash concurrency, explicit `user_id` isolation rules, and request keep-alive handling.
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.
New models, availability, and price resets that change what you run and what it costs.
Kimi K3.1 Pro and DeepSeek V4.1 Flash both target agent workloads, and they lower the cost per step in different ways. Kimi keeps one flagship model and sells a Highspeed tier that Moonshot says runs 2 to 3 times faster at twice the token price. DeepSeek builds a separate lightweight variant with a small active path, a $0.006 cached-input rate and half-price off-peak hours. Long, multimodal runs fit Kimi; high-volume loops with repeated prefixes fit DeepSeek.
Kimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index, fourth worldwide behind Claude Opus 5.5, Claude Sonnet 5.5 and Meta's Muse Wonder Full, and two points ahead of GPT-6 Astra and Gemini 4 Argon. It has the same 2.8T total and 104B active parameters as K3; the 11 point gain came from post-training. It leads the independent Terminal-Bench 4.0 run at 67.2% and costs about $2 input and $10 output per million tokens, the same tier as Claude Sonnet 5.5. Weights are promised before November 4. Human voters still rank it eighth.
Kimi K3.1 Pro is live, with a 1-million-token context window and API prices of $2 for input and $10 for output per million tokens. Highspeed doubles those input and output prices. Kimi promises model weights by November 4; its performance claims still need independent testing.
Moonshot AI released Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed on October 23, 2026. The company describes Pro as a 3-trillion-parameter-scale model with a 1-million-token context window. Both versions are available through Kimi's apps, Kimi Code and its API. Highspeed is advertised as roughly 2 to 3 times faster at output and costs twice as much per token. Full model weights are promised by November 4; they were not part of the October 23 launch.
By October 2026, products such as Muse and Grok Bot made personal agents a concrete proposition: assistants that can keep working after you close the app. The defining feature is persistent work, backed by saved task state, tools, and schedules or event triggers. A dedicated cloud computer or separate identity can help, but neither is universal. This guide explains what these agents can do and what to check before giving one access to your accounts.
A Reddit creator gave DeepSeek V4.1 Flash 1017 and two Claude models the same JavaScript room-building task. Two comparison videos, followed by a separate harness test, raise a useful question: how much of a coding demo reflects the model, and how much reflects the tools around it?
DeepSeek V4.1 Flash 1017 (max) scores 45 in the charts, up 6 points from the previous version, while cost per task falls from $0.49 to $0.26. MiniMax M3.1 Flash scores 42 at $0.22 per task.
MiniMax M3.1 Flash brings competitive business automation results and flat API pricing. DeepSeek V4.1 Flash-1017 leads on the reported coding tests and charges less for cached input. The better choice depends on the work you need an agent to finish.
MiniMax's October 19 teaser suggests Space Bunny is very likely connected to the company. A new model release may be approaching, though the exact model version remains unconfirmed.
DeepSeek V4.1-Flash-1017 improves on the September Flash release in the company's coding benchmarks. SaveMyToken's third-party roundup finds capable refactoring and prototyping, but weaker long-log retrieval, terminal discipline, and interface polish.
DeepSeek V4.1-Flash-1017 is live on the web, mobile app, and API. DeepSeek reports stronger coding and agent performance, shorter task runtimes, and less output, with API prices unchanged. Its release demos show complete projects built from brief requests. Developers can keep using deepseek-flash and should update DeepSeek Harness to 0.2.1-rc.1 or later.
Kuaishou's Kling AI switched on Kling 4.0 in the early hours of October 16, 2026, about two and a half weeks after Kling 4.0 Flash opened for early access. The headline capability is native 30-second generation in a single uninterrupted take, with up to 10 keyframes of reference control. Launch-night tests from the Chinese AI blog Digital Life Kazk show it holding a chase, a circus one-take, and fight choreography without falling apart. A 30-second 1080p clip costs 450 credits, roughly ¥16 at member pricing; 720p lands near ¥9. A 10-bit 4K HDR mode was announced but is not live yet.
Step 5 Preview is a sparse mixture-of-experts model with 600B total parameters, 27B active parameters per token, and up to 1M tokens of context. StepFun made the model weights public on October 15, after launching the preview and API on September 20. In a hands-on test by Chinese independent developer Lin, it turned a UI mockup and a short product brief into a working Agent dashboard, then fixed a preview-panel bug found during manual testing.
DeepSeek's official status page records a web and API incident starting at 08:11 Beijing time on October 15. The V4 Pro API, V4.1 Flash API and chat service were affected. At the 12:33 Beijing time check, DeepSeek had identified the cause and was implementing a fix; the incident was not marked resolved. There is no official evidence linking the outage to a V4.1 Pro gray test.
Meta's Muse Wonder 1.0 Standard API costs $4 per million input tokens, $20 per million output tokens, and $0.20 per million cached input tokens. DeepSeek V4.1 Flash lists lower token rates, with separate peak and off-peak prices. Muse Wonder Full has no published Model API token rate for this comparison. Prices checked October 14, 2026.
Google opened Gemini 4 Argon to paid API users on October 13. Its Standard launch rates are $2 input, $10 output, and $0.10 cached input per million tokens through November 30; DeepSeek V4.1 Flash lists lower off-peak and peak token rates. Prices checked October 14, 2026.
GLM-5.3 Highspeed has appeared in ZCode, with an early user reporting 490 output tokens per second. Pi lists the model as glm-5.3-highspeed. ZCode now documents a randomized GLM HighSpeed trial for subscribers; the reported speed and the proposed Omen Alpha connection remain unverified.
DeepSeek will cut V4 Flash off-peak API rates at 12:00 Beijing time on September 10, lowering cache-hit input from RMB 0.05 to 0.02 per million tokens, cache-miss input from RMB 1.50 to 1.00, and output from RMB 4.50 to 4.00. Peak rates remain twice the off-peak price.
A temporary DeepSeek V4.1 Flash route is available for limited testing through the existing API base URL. The reported beta keeps V4 Flash billing, caps each account at 20 concurrent requests, and uses a model ID that points to a September 10 expiry.
A hands-on test asks DeepSeek, ChatGPT, and Gemini to recognize the same image and turn Girl with a Pearl Earring into a Perler bead pattern, with the DeepSeek and GPT results shown side by side.
DeepSeek has launched the experimental V4 Flash Vision Exp model on its API, with mixed image and text input, three request formats, and a free Files API for reusable uploads.
DeepSeek Harness RC.8 adds native image-request configuration for DeepSeek adapters, image input for core commands, and on-demand Claude Code and Codex subagent bundles. The official release also improves search concurrency, Windows PowerShell sessions, storage, and the Python SDK.
DeepSeek will introduce separate peak and off-peak API pricing for V4 Flash and V4 Pro on August 17, 2026. This chart compares the old rates with the new prices for cache-hit input, cache-miss input, and output tokens.
DeepSeek V4 Pro 0813 went live at about 11 p.m. Beijing time on August 13. The model keeps the same API name and current rates, but DeepSeek is warning users to expect a significant price increase soon.
DeepSeek V4 Pro 0813 is now official. On August 17, DeepSeek will introduce peak and off-peak API pricing, with V4 Pro rates rising by as much as 12 times the previous price.
Updated August 13: DeepSeek-V4-Pro-0813 is now live. This July 31 release introduced the official Flash build, its large agent gains, and native Responses API support for Codex.
Updated August 13: DeepSeek V4 Pro 0813 is now live, largely confirming the reported mid-August model window. DeepSeek Harness remains pending.
Updated August 13: DeepSeek V4 Pro 0813 is now live through the existing API model name. The earlier gray-rollout signals are now historical context; DeepSeek Harness remains pending.
Moonshot AI's Kimi K3 scores 57 on the current Artificial Analysis Intelligence Index, ahead of Claude Opus 4.8, GPT-5.5, and DeepSeek V4 Pro. Its full weights are promised for July 27, and its API price is 30% of Claude Fable 5's, not one eighth.
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 June 30, 2026: Chinese financial and technology outlets report that DeepSeek upgrade emails point to a mid-July V4 formal release and a peak-valley API pricing policy. Current public DeepSeek docs still show the flat V4 Pro and Flash price table, so this item treats the peak-hour policy as a reported pricing update rather than a storefront or inventory change.
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 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 18, 2026: DeepSeek's official research and GitHub surfaces already include multimodal work through Janus-Pro and DeepSeek-VL2, covering image understanding, OCR, document/chart reasoning, visual grounding, and text-to-image generation. Official V4 API docs still list V4 Pro and V4 Flash as text-first API routes, and DeepSeek's Copilot page says V4 itself is text-only with an optional vision proxy, so this should be framed as DeepSeek's multimodal ecosystem expanding rather than a newly confirmed native V4 Vision API.
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 9, 2026: DeepSeek's English homepage still links directly to the April 24 @deepseek_ai V4 Preview post, while current official docs keep V4 Pro and V4 Flash as the active model names and mark legacy aliases for July 24 retirement.
A stronger community DeepSeek V4 Flash GGUF page now ties named files to direct llama.cpp, Ollama, vLLM, and Metal-backed reproducibility notes, while the TUI and Claude Code tracks stay unchanged enough to skip new public claims.
GitHub Releases now marks CodeWhale v0.8.43 as latest, keeps `codewhale` canonical, and leaves the DeepSeek local-deployment and Claude Code tracks unchanged enough to skip publishing new guide claims there.
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's current pricing page now says V4-Pro API pricing will be officially adjusted to one quarter of the original price after the 75% discount ends on May 31, 2026. Treat this as an official price reset, not a guarantee that prices can never change.
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.
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.
A new report says DeepSeek is preparing a June V4.1 release with stronger MCP support, enterprise tooling, and image/audio input that still returns text. The same funding cycle points to a roughly $45B-$51.5B valuation range, with state-backed capital and founder Liang Wenfeng expected to play central roles.
DeepSeek V4 is text-only in the public API today. The Visual Primitives report shows DeepSeek's vision research, but not a generally available vision endpoint.
Community testing confirms that DeepSeek V4's context caching delivers 90%+ hit rates in real workloads. With nine out of ten input tokens now billed at one-tenth the original cache price, the effective blended input cost lands around 80% below the pre-cut baseline — far more than the headline price cut alone would suggest.
DeepSeek has turned pricing into the top story: V4 Pro now sits around $0.90 equivalent in headline pricing, while cache-hit input drops to RMB 0.025 per 1M tokens. For repeated coding and agent workloads, the effective bill can fall much lower than the headline rate.
Fresh community work lowers the first credible Mac target from only ultra-high-memory experiments to a narrower 128GB path: a new ~90 GB antirez IQ2XXS quant, a dedicated fork, and a reproducible bootstrap script. This remains a community workflow, not official Mac support.
Community work around DeepSeek V4 Flash has moved from screenshots into reproducible local-deployment notes: GGUF packaging, compatible llama.cpp branches, and Mac memory limits are now the key watch points.
Independent model trackers are beginning to matter more than launch claims because DeepSeek V4 now has two different routes: Pro for harder reasoning and Flash for high-volume cost-performance.
DeepSeek's April 24 official release makes V4 the new flagship family with 1M context, named Pro and Flash variants, and pricing that has since been cut further. V4-Pro is now listed at $0.435/$0.87 per 1M input/output tokens, with those rates becoming the official quarter-price baseline after the 75% discount window ends on May 31, 2026.
Cross-platform chatter on Twitter, Reddit and Hacker News converges on three claims about DeepSeek V4: a 2,000,000-token context window, a dynamic sparse MoE architecture, and end-to-end pre-training on Huawei Ascend.
DeepSeek has officially archived the DeepSeek-V3 model with a v1.0.0 release, primarily for generating a permanent DOI.
Claude Code, TUI, local deployment, and agent workflow updates. Not required for existing API calls.
Both are top flagship models, but they structure speed and cost tradeoffs differently. Moonshot's Kimi K3.1 Pro offers one underlying model with two serving tiers: Moonshot says Highspeed delivers 2 to 3 times the output speed at double the token price. Gemini 4 Argon ships as a single flagship endpoint, with latency and cost tuned via thinking levels and API service tiers. For Kimi, you pick a model ID on each request. For Argon, you set a thinking budget and service level.
Kimi K3.1 Pro and Claude Opus 5.5 both target long-running coding and Agent workflows with a 1M-token context. The key difference is packaging. Kimi is a 3T model on an open-weight roadmap, offered via a standard ID plus a separate Highspeed ID at double the token price. Claude Opus 5.5 is a closed flagship priced at $4 input / $20 output per million tokens; its fast mode lives as a request parameter on the same model. Kimi offers a planned self-hosting path and developer-controlled search endpoints. Claude delivers a consistent cross-cloud model and model-triggered cited search.
Kimi K3.1 Pro and GPT-6.1 Sol both target cheaper Agent workloads, but take opposite approaches. Moonshot maintains one 3T-class model with up to 1M-token context on premium tiers, selling speed as an extra serving tier. OpenAI releases Sol as a lighter, independently tuned model, priced at $2 input / $10 output per million tokens. Long, context-heavy sessions suit Kimi; short, high-volume loops suit Sol.
Z.ai and Tsinghua's Jie Tang team have open-sourced HarnessPro, a framework that lets a coding agent rewrite its own harness between runs. It keeps a Pareto set of three operating points: Lite for cost, Medium for balance, and Heavy for performance. On SWE-Bench Pro with GPT-5.5 and Codex, Medium reached 79.0% pass rate at $1.26 per task, 53.0% below RHO's $2.68. The trade-off is explicit: Medium gives up 1.0 percentage point against RHO. Heavy beat RHO on all three benchmarks while costing less.
Early reports describe Kimi K3.1 building and revising Three.js and WebGL scenes, with promising static composition and a reported weakness in character animation. The claimed 1/14 cost versus GPT-6 Astra and 70% of Kimi K3 cost are tester reports without a published evaluation method. The official Kimi API model list checked on October 23, 2026 does not list K3.1. Treat the demos as evaluation leads, not a production recommendation.
DeepSeek open-sourced DeepInfer on October 20, 2026. The release provides offline inference for DeepSeek-V4.1-Flash on NVIDIA SM100 and Huawei Ascend 950. Its value is a concrete reference for model and engine co-design: architectural KV reuse, bounded replay, and shared model logic across two hardware paths. A complete online service still needs cache storage, disaggregated deployment, and load balancing. The chip-market implications below are editorial analysis, not a cross-platform performance test.
DeepSeek shipped version 0.2.1-alpha.4 of DeepSeek Harness (DSH) on October 15, 2026. Two features lead the release: reasoning levels for individual custom API models can now be set in the UI instead of by editing configuration files, and DSH can discover and install plugins from custom plugin index sources. Five bug fixes cover command input, failed-command status, Office/PDF file delivery, the Developer Tools panel, and Linux startup. One change matters to plugin authors: engines.dsh now takes priority when a plugin declares it. DSH remains a pre-release developer preview, so expect breaking changes. If you are already on the alpha line, this patch is worth taking.
DeepSeek released DeepInfer on October 20 alongside the deepseek-recipe codec libraries. DeepInfer exposes the inference engine behind V4.1 Flash, including distributed execution, DSpark speculative decoding, SWA bounded replay and asynchronous scheduling. Recipe handles conversion between API protocols and model tokens. The initial repository provides offline inference; a complete online service still needs distributed KV storage, deployment across separate serving stages and load balancing across replicas. The release shows how model architecture, cache lifetime and hardware scheduling can be optimized together.
Connect DeepSeek through the built-in DeepSeek provider in Hermes. Then confirm the model and API endpoint, create separate profiles for different tasks, and enable only the tools you need. Test the setup on small tasks with clear pass criteria. If results deteriorate or responses take too long, check the model version, reasoning settings, conversation history, and tool responses one at a time.
DeepSeek Harness (DSH) alpha.4 (release tag dsh-v0.2.1-alpha.4, published October 15, 2026 at 21:40 Beijing time) can now discover and install plugins from custom index sources instead of only the official catalog. One hour and 48 minutes after the release, developer Alexander Kloster had all 90 agents from the Grok Bot marketplace packaged as DSH plugin bundles and serving from a live index, at 23:28 the same evening. The presets can be selected in DSH after installation. Installation takes four steps. The index is community-run, the repo carries no license file, and this is alpha software. Don't hand these bots your API keys.
A developer going by FeiZ has open-sourced DeepSeek Bot, a Grok Bot alternative that runs as one plugin in DeepSeek Harness: tell the supervising bot Chief what you need, and each teammate gets its own name, model, and long-term memory.
DeepSeek Harness v0.2 is now in preview, with desktop installers for macOS and Windows, built-in tools for everyday work and development, and a new plugin management page.
DeepSeek's DSec infrastructure supports stateful agent training with four sandbox backends, composable environments, on-demand image loading, and rollout recovery. Read our technical analysis and a full English translation of DeepSeek's Chinese article.
DeepSeek Harness v0.1.2-alpha.3 improves long-session navigation and rendering, fixes queued image delivery, and removes the optional SQLite Session backend. The alpha remains separate from npm's default RC release.
Some DeepSeek users say they are seeing an earlier-looking reasoning style and better long-task results in V4 Pro. The reports may point to a traffic experiment, but DeepSeek has not confirmed a new model, checkpoint, or routing change.
DeepSeek has released its Harness developer preview under the MIT license. The agent runtime makes the model adapter, tools, session log, and even the agent loop replaceable through plugins.
Checked on July 13, 2026: DeepSeek's official docs sitemap currently lists both a top-level Prompt Library page and an Integrate with AI Tools guide covering Claude Code, OpenCode, and OpenClaw, giving developers a first-party docs path from prompt exploration to concrete tool configuration.
Checked on July 8, 2026: DeepSeek's official docs sitemap currently lists dedicated thinking-mode sample pages for both non-streaming and streaming chat completions, clarifying when `reasoning_content` is just captured, when it is ignored in ordinary follow-up turns, and how `delta.reasoning_content` should be accumulated separately from visible answer text.
Checked on July 2, 2026: OpenAI's Codex docs now document custom model providers through `model_provider`, `model_providers`, `base_url`, and `env_key`. That means DeepSeek can be framed as a third-party Codex route: use our DeepSeek access, point Codex at the configured DeepSeek provider, and keep the known Responses API, profile, Desktop picker, and V4 thinking-mode checks in place.
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.
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 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 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 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.
Checked on June 10, 2026: DeepSeek's official Claude Code page says the DeepSeek API natively supports Web Search in Claude Code, while the same vendor page still pins the main model to V4 Pro and subagents to V4 Flash.
The maintained public CodeWhale surfaces now show v0.8.50 as the stable baseline, keep `codewhale` as canonical, and reinforce `.codewhale` plus checksum-verified release artifacts as the setup path.
DeepSeek TUI has moved to CodeWhale v0.8.49 with a newer release/install surface, and DeepSeek's own Anthropic docs now spell out Claude-style model mapping for Claude Desktop developer mode.
CodeWhale v0.8.47 is now latest, while the current README and Docker docs switch the primary config and volume path to `.codewhale` with `.deepseek` kept only as a legacy fallback.
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.
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.
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.
Two Chinese developer reports point to the same product signal: DeepSeek is hiring for Agent Harness roles and, according to sources cited by Jiaziguangnian plus public remarks from researcher Chen Deli, is organizing a Harness effort around code-agent products comparable to Claude Code.
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.
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`.
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.
The post-launch DeepSeek V4 story is moving toward deployment economics: long context, Flash pricing, open weights, and hardware strategy are now the angles that matter for developers and buyers.
Third-party explainer sites deepseekv4.dev and deepseek-v4.ai aggregate the leaked spec sheet: ~1T-parameter MoE, Engram memory core, mHC hyper-connections reasoning, and a 1M+ token context window.
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.