Actualités DeepSeek

Actualités DeepSeek

Suivez les actualités des modèles DeepSeek, de l’API, des tarifs et des outils pour développeurs, ainsi que les nouvelles de l’IA. DeepSeek V4 Hub est une publication indépendante.

Dernières actualités

135
OfficialArticle en anglais

DeepSeek Harness v0.2.1-alpha.4: UI Reasoning Tiers and Custom Plugin Indexes (Oct 2026)

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.

Lire l’article
OfficialArticle en anglais

DeepSeek open-sources DeepInfer inference framework and DeepSeek-recipe codecs, benefiting third-party deployments

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.

Lire l’article
CommunityArticle en anglais

DeepSeek not working well in Hermes? A practical guide to setup and troubleshooting

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.

Lire l’article
CommunityArticle en anglais

90 Grok Bots Move to DeepSeek Harness: alpha.4 Plugin Markets (Oct 2026)

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.

Lire l’article
OfficialArticle en anglais

DeepSeek launches V4.1-Flash-1017: one-click coding brings ideas to life, closing in on GPT-5.6

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.

Lire l’article
AI VideoArticle en anglais

Kling 4.0 Drops: 30-Second Continuous One-Take Video. The Kling We Know Is Back

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.

Lire l’article
Open WeightsArticle en anglais

Step 5 Preview: China's New Open-Weight 600B Sparse MoE with a 1M-Token Context Window

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.

Lire l’article