DeepSeek V4.1-Flash-1017 Adds Live Web Search and Beats Claude Opus 5 on Coding

What Shipped on October 17

DeepSeek pushed a new checkpoint of its V4.1-Flash model live on October 17, 2026 — DeepSeek-V4.1-Flash-1017 — and made it available on the company's website, app, and API the same day. The update is served under the deepseek-flash model name and published to Hugging Face under an MIT license, keeping the line's open-weights tradition intact.

DeepSeek claims the new checkpoint beats the September V4.1-Flash release on all 11 benchmarks the company reported, with gains ranging from 2.0 to 11.1 points — alongside faster task completion and shorter outputs. The headline result: a 77.0 score on DeepSWE v1.1, a real-world software engineering benchmark, which DeepSeek's own table places ahead of Claude Opus 5's 74.0.

A Model Release With No Press Release

Part of the story is how the release happened. OpenAI, Anthropic, and Google treat a model update as an event — a blog post, a model card, usually a livestream. DeepSeek treats it as a config change. The first public sighting of the 1017 checkpoint came from bot accounts that poll the API and diff the model list, with community trackers catching the name appear and briefly disappear before the full rollout landed.

For developers watching the API, the practical takeaway is simple: if you're calling deepseek-flash, you may already be running the new checkpoint. There's no migration step, no new endpoint, and no pricing change — the dated snapshot simply replaced what came before, consistent with DeepSeek's habit of numbered releases like the earlier V4-Flash-0731.

The Benchmarks: 77.0 on DeepSWE

The DeepSWE v1.1 result is the one that matters most for anyone using AI coding tools. The benchmark measures whether an agent can resolve real GitHub issues end-to-end — understanding a repository, locating the relevant code, and producing a working fix. DeepSeek's 77.0, ahead of Claude Opus 5's 74.0 on the same table, suggests the gap between the cheapest frontier-tier API and the most expensive ones is now measured in single digits on agentic coding workloads.

The picture isn't uniformly dominant. On HLE (Humanity's Last Exam) without tools, the new checkpoint lands at 42.6 against Opus 5's 56.3 — a reminder that deep reasoning without retrieval or code execution remains the domain of the biggest closed models. DeepSeek is explicitly pitching this release as a straight upgrade on cost and speed for agentic work rather than a frontier-reasoning tradeoff.

The Real News: Search and Fetch APIs Go Public

Bundled with the checkpoint is something arguably bigger than the model itself: DeepSeek opened its Search and Fetch APIs to all developers, no longer restricted to internal use. The Search API gives applications live web retrieval, priced at $0.75 per 1,000 queries, while the Fetch API pulls page content for grounding.

This matters because live retrieval has been the missing piece for cheap, self-hostable agent stacks. Open-weight models are free to download, but an agent that can't search the web is stuck with a training-cutoff view of the world. With an MIT-licensed 1M-context model and a low-cost retrieval API from the same vendor, a complete open agent stack — model, context, and live data — now exists at a fraction of frontier pricing. Expect the AI coding assistants, research tools, and chatbot builders in the ecosystem to add DeepSeek search grounding as an option in the coming weeks.

Inside the 552B MoE Architecture

The V4.1-Flash line that the 1017 checkpoint updates is architecturally interesting in its own right:

• 552B backbone, tiny active compute: A sparse Mixture-of-Experts that activates roughly 8B parameters per token on input and 16B on output — prompt-heavy workloads (long documents, retrieval context, codebases) run on the cheap half of the model
• 384 routed experts + 1 shared expert per MoE layer, with 6 routed experts active per token
• 1M-token context window with output capped at 384K
• Native multimodality: text and image understanding built in from pretraining, trained on 45T tokens — the separate vision variant is gone
• 196B Engram conditional memory sparsely accessed via token-based lookup, on top of the backbone
• MIT license with vLLM and SGLang deployment recipes — weights are ungated on Hugging Face

The full checkpoint, including memory tables, totals around 763B parameters and roughly 510 GB on disk — so "open" doesn't mean "runs on your laptop." But for teams with GPU capacity or API budgets, the economics are hard to argue with.

Pricing and Availability

V4.1-Flash pricing carries over to the 1017 checkpoint. Peak rates run $0.30 per million input tokens and $1.20 per million output tokens, with off-peak at half that and cache hits down to $0.003 per million. The Search API is priced separately at $0.75 per 1,000 queries.

The model is live on the DeepSeek website and app, through the API at api.deepseek.com (OpenAI-compatible format), on Hugging Face for self-hosting, and through gateways like Vercel's AI Gateway, which already lists the October 17 snapshot.

Pros and Cons

What This Means for AI Tool Users

If you're a developer, this is the strongest cost-to-performance ratio in the open-weight ecosystem right now, and the newly public Search API removes the last excuse for building agents on stale knowledge.

If you're choosing AI tools, expect the coding assistants and research agents you already use to quietly get better and cheaper as they route more workloads to models like this — the same way Vercel's gateway picked up the snapshot within hours.

If you're building a product, MIT weights plus a $0.75-per-thousand-queries retrieval API means a complete agent stack with no vendor lock-in is now a realistic architecture, not a science project.

You can compare DeepSeek-powered tools and hundreds of alternatives on aitrove.ai.

Frequently Asked Questions