DeepSeek pairs Flash with web tools

DeepSeek has released new model weights for its cost-efficient coding model and made its Search and Fetch APIs available to all API accounts. Developers can now combine a long-context model with live web retrieval under the same account, authentication system, and pricing structure.

The October 17, 2026 release, named DeepSeek-V4.1-Flash-1017, updates the V4.1 Flash family while retaining the hosted API identifier deepseek-flash. DeepSeek reports higher benchmark scores, faster task completion, and shorter outputs than the previous snapshot at the same token prices.

Flash spends compute unevenly

V4.1 Flash uses a sparse mixture-of-experts architecture with 552 billion total parameters. Input processing activates 8 billion parameters, while output generation activates 16 billion. Activating only a fraction of the model for each token lowers compute requirements while preserving access to the larger parameter pool.

The technical report describes a 40-layer Causal Encoder-Decoder network divided evenly between encoder and decoder layers. Its feed-forward layers use DeepSeekMoE. The first two encoder layers use sliding-window attention, while the remaining layers use Compressed Sparse Attention 2 to select a learned subset of earlier tokens instead of attending to the entire history.

Compressed key-value caching reduces cache memory to about one-quarter of the previous Flash generation, according to DeepSeek. The key-value cache stores attention state from earlier tokens during generation and can consume much of the available GPU memory in long coding sessions or repeated tool-use loops.

Built for long agent runs

DeepSeek designed the snapshot for coding, terminal, computer-use, and research agents that operate across many steps. Its training included reinforcement learning with self-feedback for end-to-end project generation, including games and 3D models. Image data was included during pretraining, giving the model native image-understanding capabilities.

DeepSeek also reports that Flash exceeds V4 Pro in benchmark performance, generation speed, and total task-completion time. Those comparisons are vendor-reported, so application-level testing remains necessary for specific repositories, tools, and agent loops.

Search and Fetch open to every account

DeepSeek has also made two JSON APIs generally available. The Search guide describes an endpoint that returns structured web results for a query, while Fetch retrieves one page and converts it to Markdown. Both services use the same bearer-token authentication as the chat API.

Fields that retrieval pipelines can use

• Search forwards operators including site:, -site:, lang:, after:, and before:.
• Each result includes a rerank_score from 0 to 3 and an authority tier from 0 to 4 for filtering and ranking.
• Empty searches and failed fetches are excluded from billing.
• Fetch returns Markdown, reducing the boilerplate removal normally required before indexing web pages for retrieval-augmented generation.

Minimal Search request

import os
import requests

response = requests.post(
    "https://api.deepseek.com/v1/tools/search",
    headers={
        "Authorization": f"Bearer {os.environ['DEEPSEEK_API_KEY']}"
    },
    json={
        "query": "DeepSeek V4 release lang:en after:2026",
        "count": 5,
        "need_content": True,
    },
    timeout=30,
)
response.raise_for_status()

for item in response.json()["results"]:
    print(item["rerank_score"], item["url"])

DeepSeek says Search and Fetch calls initiated through DeepSeek Harness, Claude Code, and Codex are free for existing API users until billing begins on October 31, 2026. The company also says the underlying search system already handles hundreds of millions of internal requests each day.

The cost model for agents changes

Long-running agents accumulate costs from model inference, retained context, and external tool calls. Flash targets the first two through low token prices and a smaller attention cache, while Search and Fetch price web access at $0.00075 per successful request. One million successful web requests would cost $750 at the listed rate.

The account-level limit of 5 requests per second permits up to 18,000 requests per hour when sustained, which may constrain parallel agent fleets. Production evaluations should measure contention at that limit alongside search relevance, fetch quality, output-token use, cache memory, and end-to-end completion time.

MIT-licensed weights allow teams to begin with DeepSeek’s hosted model and move inference to their own infrastructure later. Search and Fetch remain managed DeepSeek services, so self-hosted inference still carries an external retrieval dependency unless the application replaces that layer.