DeepSeek Harness gains enhanced support for V4.1-Flash

DeepSeek has released V4.1-Flash-1017 and updated DeepSeek Harness to support it. The update also addresses model routing, timezone management, and web search.

DeepSeek has launched DeepSeek-V4.1-Flash-1017 and simultaneously pushed an update to DeepSeek Harness, the tool that serves as an interface and orchestrator for multiple platform uses. The goal is clear: better support the new model while adjusting very concrete components like Git branches, time zones, and session display. All this arrives as the new set is presented as faster, more efficient, and less verbose than the previous version.

Un modèle plus rapide, et des outils qui suivent le rythme

On X, DeepSeek describes V4.1-Flash-1017 as a “Pareto improvement in performance and cost,” meaning a simultaneous gain in performance and cost. The company also states that the model produces shorter outputs and completes tasks faster than the previous generation. This is not just a changelog line: DeepSeek says it has already deployed it on its website, app, and API.

In the same message, the company specifies that the model is optimized for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. The recommendation is explicit: upgrade to DeepSeek Harness v0.2.1-rc.1 or later to fully benefit from the new model. A polite way of saying that without an update, the mechanism may start to creak.

Ce que DeepSeek Harness change vraiment sous le capot

The technical notes list several changes on the harness side, with a mix of localization, interface, search, and model routing. Japanese support is enhanced, with a language pack, a 60-entry dictionary for namespaces, and less rigid CJK font management. The idea is simple: let the user see Japanese properly rendered without imposing a single font on everyone.

The platform also improves Git branch reading in the composer, workspace name display in session lines, and message time presentation. The timestamp is no longer stuck in the action line: it moves below the message, with a configurable timezone in settings, from terminal to JST or UTC. It seems like a detail, but teams juggling long sessions know that a wrong time can quickly turn a history into a detective novel.

Pourquoi le routage des modèles prend de l’ampleur

Another part of the update concerns routing to multiple providers, including Ollama Cloud and OpenCode Go. DeepSeek Harness can now differentiate availability windows by local hours, weekdays, and associated IANA timezone. Practically, the system can switch a model to another based on time slot, expected load, or configured profile.

The logic is designed for production use. The example given in the notes shows a split between a daytime provider and another at night, with priority on routes already configured in the profile. The selected model remains visible in the interface, with the provider’s name and, if applicable, an automatic failover mention. The front end thus displays what the infrastructure decides behind the scenes, avoiding surprises when everyone thinks they are using the same component.

Recherche, sessions et fuseaux : des détails qui pèsent lourd

DeepSeek also adds a web search backend via POST /api/web_search, authenticated with OLLAMA_CLOUD_API_KEY. The service returns sources but no generated response, and it can exclude entries without URLs. On its side, the model selection in the interface shows the provider, model, and effort level, with automatic tagging during peak windows.

Sessions themselves evolve: the working directory can change without breaking replay history, and folder changes are preserved as long as the original project still exists. Finally, sending session logs to the official DeepSeek API remains opt-in, disabled by default. The original data does not change, but its routing is now finely tuned. It’s the kind of detail that goes unnoticed until the day it saves a team from a bad setting.

By the numbers

  • v0.2.1-rc.1 — minimum recommended version of DeepSeek Harness, according to DeepSeek.
  • 60 entries — Japanese dictionary added for namespaces.
  • $0.75 per 1,000 calls — announced price for each Search & Fetch endpoint.
  • 4.4M views — volume displayed on DeepSeek’s X post, October 17, 2026.
  • 2026 — publication date visible on the source post shared by DeepSeek.

To read