DeepSeek news & updates
One ongoing page for DeepSeek models, API changes and open-weight releases. Verified developments are added to the timeline below, with the sources and context needed to follow what changes.
1 substantive entry added. Coverage: 17 October 2026 at 12:08 BST–18 October 2026 at 12:09 BST. All historical entries and existing visuals remain on this page. Next scheduled scan: 19 October 2026 at 12:00 BST · Europe/London.
The ongoing timeline
Newest update firstHistory stays here
Perplexity lists a hosted DeepSeek V4 Pro 0813 endpoint
Perplexity now documents a hosted DeepSeek V4 Pro 0813 endpoint for Agent API and Orchestrator API. The endpoint is run by Perplexity; this announcement is not a new DeepSeek checkpoint, weight release or change to DeepSeek’s own API.
Why it matters: It adds another documented route to an existing model, with Perplexity’s hosting and billing.
Perplexity’s official RSS dates the notice to 17 October 2026 at 12:16 BST. The first callable moment is not separately specified.
DeepSeek Harness release candidate adds Flash-1017 and web-retrieval controls
DeepSeek Harness 0.2.1-rc.2 adds support for DeepSeek-V4.1-Flash-1017 and brings Search and Fetch into its Web Search settings. It follows the first 0.2.1 release candidate and the earlier alpha series already recorded here. GitHub still marks the package as a prerelease, so these changes are not a stable-channel launch; the model weights and separately billed retrieval APIs are covered in their own entry below.
Why it matters: The agent runtime can expose the newly released model and retrieval tools in one workflow, while release-candidate status and API billing remain separate considerations.
Release candidate 1 was published on 17 October 2026 at 06:22 BST; rc.2 followed at 09:16 BST (08:16 UTC). This materially updates the existing Harness event without changing its original anchor or 10 October added time.
Updated 17 October 2026 at 12:08 BST: Added the release-candidate follow-up. Earlier alpha coverage, sources and the original added time are retained below.
Previous published coverage — historical record, not current availability
DeepSeek Harness alpha.4 adds custom-model reasoning controls and plugin-index sources
DeepSeek Harness 0.2.1-alpha.4 adds UI controls for reasoning levels on individual custom-API models and lets users discover and install plugins from custom plugin-index sources. It changes third-party plugin compatibility checks to prioritise declared engines.dsh ranges, and fixes command input, Office-file delivery, developer-tools activation, Linux startup and task-status issues. This remains an explicit prerelease of the MIT-licensed agent runtime, not a stable release, new DeepSeek model, API upgrade or model-weight release; the earlier alpha.3 and alpha.2 coverage is retained below.
Why it matters: Configuration and plugin discovery become more accessible, while plugin developers have a concrete compatibility change to check before adoption.
Official pre-release published 15 October 2026 at 14:40:02 BST (13:40:02 UTC). This updates the existing Harness event while retaining its original alpha.2 anchor and added time, plus the alpha.3 history.
Updated 16 October 2026 at 12:17 BST: Added the material follow-up. The original anchor, added time and earlier published coverage are retained below.
Previous published coverage — historical record, not current availability
DeepSeek Harness alpha.3 adds image-aware compaction and resumable SDK sessions
DeepSeek Harness 0.2.1-alpha.3 adds image reading and automatic long-context compaction to Minimal mode, plus a compact command for Web and Desktop. It also adds per-subagent Agent Presets, optional persistent Node execution in PTC mode, sent-secret masking and session resume in the TypeScript and Python SDKs. Legacy azure-openai-responses provider configurations need migration to azure. This is explicitly a prerelease of the agent runtime, not stable general availability or a new DeepSeek model; the earlier worktree and SSH coverage is retained below.
Why it matters: The update materially expands multimodal agent and integration workflows, while its preview status and compatibility changes matter before adoption.
Official alpha.3 release published 14 October 2026 at 14:32 BST. This updates the alpha.2 event while retaining its original 10 October added time.
Updated 15 October 2026 at 12:14 BST: Added the alpha.3 follow-up and compatibility notes. The original event anchor, added time and alpha.2 record are retained.
Earlier release coverage — retained for the record
DeepSeek Harness alpha adds opt-in worktrees and remote SSH helpers
DeepSeek Harness v0.2.1-alpha.2 adds experimental plugins for Git worktrees and translating displayed reasoning, both enabled through the plugin settings. Standalone SSH helper runtimes support remote operations, while TypeScript and Python integrations gain session working-directory controls. Custom integrations need to review compatibility changes to tool presentation, subagent execution and Agent Team messaging. This is an alpha developer-tool release, not a new DeepSeek model.
Why it matters: Isolated checkouts and remote execution expand agent workflows, but the pre-release status and integration migrations matter before adoption.
DeepSeek releases Flash-1017 weights and separate Search and Fetch APIs
DeepSeek has published V4.1-Flash-1017 under the MIT licence and updated its web, app and existing deepseek-flash API alias, without changing the model API’s pricing or calling method. The company reports better reasoning, coding, agent and visual benchmark results with shorter execution and output, but those are vendor-run evaluations that depend on the harness. Separately, its read-only Search and Fetch JSON APIs are priced at US$0.75 per 1,000 successful result-bearing requests each, with independent account-level limits of five requests per second. Server-side calls through Harness, Claude Code or Codex are temporarily free until 1 November at 00:00 Beijing time; that exception is not a blanket free offer for all API usage.
Why it matters: The release combines downloadable model weights, an updated hosted model and a separately metered retrieval layer, which have different operational and billing implications.
Official model and API announcements are dated 17 October 2026. The model repository was created that morning; failed, empty-result and timed-out retrieval requests are not billed under the published guide.
| Route | What the source establishes |
|---|---|
| Downloaded checkpoint | MIT-licensed model weights; running infrastructure is the user’s responsibility. |
| Hosted deepseek-flash | Existing alias updated; published model API pricing is unchanged. |
| Search / Fetch | Separate per-request tools: US$0.75 per 1,000 successful requests, with the limited promotion described above. |
DeepSeek adds sparse-attention operators to its Ascend kernel library
DeepSeek merged its October public source update for DeepGEMM-Ascend, advancing the library’s source version to 0.1.1. The revision adds sparse and paged-sparse multi-query-attention operators and related Python bindings, alongside dense matrix-multiplication bias support and BF16 stochastic rounding. The MIT-licensed project targets Huawei Ascend hardware, and this is a source-code update on the main branch, not a tagged package release, new DeepSeek model or newly available model weights. No independent workload benchmark is established by the release record.
Why it matters: Low-level operators help developers adapt training and inference to hardware; source availability and additional operations do not prove faster application performance.
October public source revision merged to main on 14 October 2026 at 10:22:30 BST; source version 0.1.1.
DeepSeek V4.1 Flash adds native vision and a new architecture
DeepSeek released V4.1 Flash with native visual understanding and a new asymmetric architecture intended to reduce inference and context storage costs. The model is available through the DeepSeek API as deepseek-flash, with downloadable weights published under an MIT license on Hugging Face. Efficiency and benchmark comparisons in the announcement are DeepSeek’s reported results.
Why it matters: Native vision expands the tasks the model can handle, while reduced context storage requirements may benefit long agent workflows.