DeepSeek Releases V4.1-Flash-1017 With MIT Weights and Search APIs
According to DeepSeek, the dated model update delivers higher coding scores and shorter output lengths, with the MIT open-source license applying to the model weights, whereas the Search and Fetch APIs are paid proprietary cloud endpoints.
In short
- DeepSeek-V4.1-Flash-1017 is an open-weight dated checkpoint released under the MIT license on Hugging Face and ModelScope.
- The model improves scores across agentic coding benchmarks, reaching 77.0 on DeepSWE v1.1 and 42.3 on Terminal-Bench 4.0 while DeepSeek reports reductions in execution time and output tokens.
- DeepSeek made its in-house Search and Fetch APIs publicly available at $0.75 per 1,000 requests, with free access via select developer tools through October 31, 2026.
- Existing API customers access the update automatically under the deepseek-flash model ID with token pricing unchanged at $0.15 per million input and $0.60 per million output tokens.
DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, 2026, rolling out an updated open-weight model across its web, mobile, and API platforms alongside standalone web search and page-fetching tools. Available under the MIT License on Hugging Face and ModelScope, the dated checkpoint targets agentic software development, reasoning, and visual tasks. The company characterized the update as a Pareto improvement, asserting that the new release delivers higher benchmark marks while simultaneously shortening generation length and execution time to reduce total operating costs.
The launch immediately transitions DeepSeek's consumer and developer services to the new model checkpoint. Existing API users calling the deepseek-flash endpoint receive the updated behavior automatically without code migrations or price adjustments, according to reporting by AI Weekly.
Benchmark Gains Across Agentic and Coding Tasks
DeepSeek's published evaluation suite compares V4.1-Flash-1017 against its previous V4.1-Flash checkpoint as well as external models including K3, GLM-5.3, GPT-5.6 Sol, and Claude Opus 5. In coding and terminal execution evaluations, the new release posted substantial gains. On Terminal-Bench 4.0, evaluated using the Claude Code harness, DeepSeek's score advanced from 31.2 to 42.3, putting it ahead of GPT-5.6 Sol (39.9) and GLM-5.3 (37.9), though behind Claude Opus 5 (51.8).
On the DeepSWE v1.1 software engineering benchmark tested via the Mini-SWE-V2 harness, the model reached 77.0, rising from 74.2 and edging past Opus 5 (74.0) and GPT-5.6 Sol (73.0). Other developer-focused benchmarks showed similar increases: ProgramBench moved from 20.3 to 29.1, NL2Repo-Bench rose from 65.4 to 68.2, Automation-Bench v1.0.6 climbed from 54.8 to 60.5, and Apex Bench v1.1 grew from 64.6 to 72.1, according to data in the Hugging Face model repository.
General reasoning and tool-assisted visual benchmarks also recorded incremental progress. On Humanity's Last Exam (HLE) without tools, Flash-1017 scored 42.6 overall and 44.0 on the text-only subset, up from 36.8 and 39.1 on the prior Flash release. Opus 5 maintained a clear lead on raw reasoning with an HLE score of 56.3. On visual benchmarks with tool access, DeepSeek logged 83.0 on Chartography and 92.5 on BabyVision. As noted by MindStudio, while frontier proprietary models retain an advantage in abstract knowledge tasks, DeepSeek's gains concentrate heavily in workflows requiring multi-step tool invocation and command-line execution.
KV Cache Compression and Prompt Formatting
The technical foundation of the release centers on memory efficiency during long-context processing. A companion technical report released alongside the checkpoint is titled "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression," as documented on DeepSeek's model card. Key-value (KV) cache compression minimizes the memory footprint needed to retain conversational history and tool transcripts, which is critical when operating across the model's supported 1 million token context window.
The model weights are packaged as an FP8 (8-bit) quantized safetensors checkpoint distributed across 48 shards. DeepSeek recommends inference settings of temperature 1.0, top_p 1.0, and a maximum output ceiling of at least 256,000 tokens for multi-turn tasks.
Unlike most open-weight models, DeepSeek-V4.1-Flash-1017 does not provide a standard Jinja chat template. Instead, DeepSeek included a standalone Python script, encoding.py, in the repository's encoding directory to handle multi-turn conversations, tool calling, numeric reasoning effort parameters, and interleaved images. For production systems, the company introduced deepseek-recipe, a suite of Rust libraries with Python bindings designed to translate Chat Completions, Messages, and Responses API formats into model token sequences and parse tool calls from streaming outputs.
Wish-Driven Programming and Agent Integration
DeepSeek highlighted an application style it terms "wish-driven programming" or "wish-style programming," where an agent constructs functional software projects from concise human prompts and minimal reference materials. According to DeepSeek API documentation, the capability stems from reinforcement learning utilizing self-generated feedback, allowing the model to critique and adjust its end-to-end outputs during training.
To demonstrate this behavior, DeepSeek published interactive project demos built in limited interaction turns, including a 3D model of the Palace Museum, a web-based racing game, a recreation of Portal 2 running on the Godot engine, and a multi-era 3D visualization called Valley of Ages, as detailed by Pandaily. For agent orchestration, DeepSeek recommended that developers pair the model with DeepSeek Harness version 0.2.1-rc.1 or later, while noting specialized tuning for developer environments including Tencent's CodeBuddy and WorkBuddy.
API Pricing and Public Search Endpoints
Alongside the model checkpoint, DeepSeek opened its proprietary search engine to external developers through new Search and Fetch APIs. DeepSeek stated that its search infrastructure handles hundreds of millions of daily queries globally. The endpoints allow external programs to query web results and retrieve page contents, priced at $0.75 per 1,000 requests per the DeepSeek API portal, with a rate limit of 5 queries per second per account reported by AI Weekly.
Server-side search calls routed through DeepSeek Harness, Claude Code, and Codex are currently free for DeepSeek API users until billing begins on October 31, 2026, at 16:00 UTC. Standard inference pricing on DeepSeek's hosted API remains $0.15 per million input tokens, $0.60 per million output tokens, and $0.003 per million cache-read tokens. In an independent 56-trial benchmark reported by AI Weekly, Flash-1017 completed a full task evaluation set for $0.0157, compared to $0.0577 for GPT-6.1 Sol and $0.0653 for Claude Sonnet 5.5.
Operational Context and Market Impact
The arrival of Flash-1017 follows a significant service interruption earlier in the week. On October 15, 2026, DeepSeek experienced an outage lasting approximately five hours, with Tencent News recording 44 total service degradation or outage incidents on DeepSeek's public status tracker in 2026, including a 12-hour disruption in late March, as reported by Pandaily and Briefs Finance. DeepSeek has not connected the outage to the deployment of the new model checkpoint.
By labeling the release a dated update to V4.1-Flash rather than a new foundation model generation, DeepSeek continues to advance performance through tuned iterations while maintaining an open MIT license and low API pricing that continues to pressure commercial competitors across the software tooling market.
Questions people ask
What license governs the DeepSeek-V4.1-Flash-1017 model weights?
DeepSeek released the model weights and associated code repository under the permissive MIT License, allowing open commercial use, modification, and self-hosting.
How can developers format prompts without a standard Jinja chat template?
DeepSeek provides a Python reference implementation in the repository's encoding folder, as well as deepseek-recipe, a Rust library with Python bindings for production formatting and response parsing.
How much do the new DeepSeek Search and Fetch APIs cost?
Both Search and Fetch endpoints cost $0.75 per 1,000 calls. Calls through server-side tools in DeepSeek Harness, Claude Code, and Codex remain free for DeepSeek API users until October 31, 2026, at 16:00 UTC.
Do existing DeepSeek API customers need to change their code to use the update?
No code changes are required. The update was applied directly to the existing deepseek-flash API model endpoint with unchanged token pricing.
Sources
- DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real | DeepSeek API Docsapi-docs.deepseek.com, official source
- deepseek-ai/DeepSeek-V4.1-Flash-1017Hugging Face, Oct 17, 2026, official source
- DeepSeek Makes V4.1-Flash-1017 Official With MIT-Licensed Weights and Stronger Agentic Codingpandaily.com, Oct 17, 2026
- DeepSeek Opens V4.1-Flash-1017 Weights and $0.75 Search APIAI Weekly, Oct 17, 2026
- DeepSeek V4.1-Flash-1017: What's New in the Benchmark UpdateMindStudio, Oct 17, 2026
- DeepSeek-V4.1-Flash-1017: Improved coding performance and speed, public Search & Fetch APIsdevotd., Oct 17, 2026
- DeepSeek V4.1-Flash-1017: Faster, Leaner AI - Briefs FinanceBriefs Finance, Oct 17, 2026
How this story was made: the newsroom picked it up from Google Search and pandaily.com, gathered the full text of the sources above, and drafted it with AI assistance. Every factual claim was then checked against those sources before publishing (33 claims checked). Illustrations marked as AI-generated are not photographs. Spotted an error? Tell us.
Published October 18, 2026 at 02:09 UTC
