huggingface.co web signal

DeepSeek Opens V4.1-Flash-1017 Weights and $0.75 Search API

By Published Updated
5 sources tracking this story
DeepSeek Open Source Inference Agents ai-business

TL;DR

  • Independent 56-trial testing puts DeepSeek's per-run cost at 3.6x below GPT-6.1 Sol on identical bounded coding tasks, at $0.016 versus $0.058 per run.
  • The update shipped under the existing deepseek-flash API identifier with no changelog; all existing consumers absorbed the behavior change before documentation appeared.
  • Weights released under MIT alongside the API launch remove the open-versus-closed tradeoff for teams pricing a switch away from OpenAI or Anthropic.

DeepSeek posted V4.1-Flash-1017 to Hugging Face on October 17, 2026 under an MIT license, pairing the weights drop with an open rate card for its Search and Fetch APIs at $0.75 per 1,000 successful requests. The model card describes the release as "a minor update to DeepSeek-V4.1-Flash, delivering substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks," and credits the update with reducing "task execution time and output length, enabling users to complete tasks faster and at lower cost."

The headline score is 77.0 on DeepSWE v1.1, up from the previous Flash's 74.2 and ahead of Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0) on that specific coding benchmark, per MarkTechPost. Terminal-Bench 4.0 jumps to 42.3 from 31.2. MarkTechPost reports the new checkpoint beats the previous Flash on all 11 benchmarks DeepSeek published, by 2.0 to 11.1 points. On raw reasoning, Opus 5 still leads: HLE comes in at 42.6 for Flash-1017 versus 56.3 for Opus.

Pricing on DeepSeek's own serve is $0.15 per million input tokens and $0.60 per million output tokens, with cache reads at $0.003/M. The new Search and Fetch endpoints are capped at 5 QPS per account; server-side calls routed through the DeepSeek Harness, Claude Code or Codex stay free until November 1, 2026. The companion paper is titled "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression," placing the memory story on the cover, and arriving in the middle of a run of tracked inference-side KV cache work over the past month.

Context is 1M tokens with max output at or above 256K, and the recommended sampling is temperature 1.0, top_p 1.0.

What others are reporting

Coverage cluster as of 25h after publish

  1. DeepSeek Read →

    Official first-party launch post confirming the 552B MoE architecture, pricing structure, and silent routing of legacy V4-Pro requests to V4.1-Flash billing rates.

    552B-parameter MoE. New Causal Encoder-Decoder architecture: just 8B active parameters for input, 16B for output.
  2. Tech Insider Read →

    Frames the release as ecosystem lock-in via low switching costs and a two-week billing grace period, rather than as a claim of benchmark supremacy over rivals.

    Gains on one axis without giving anything up on another.
  3. Kingy AI Read →

    Original 56-trial hosted testing quantifying the cost gap, with the notable admission that V4.1-Flash-1017 trails Claude Opus 5 on 7 of 11 benchmarks.

    Its October 17 update raises every score in DeepSeek's 11-row release table, keeps the existing API rates, and releases the weights under MIT.
  4. Munder Difflin Blog Read →

    Developer-facing: documents Claude Code environment variable setup and confirms deprecated model names still route silently to V4.1-Flash at the new billing rate.

    a minor update to DeepSeek-V4.1-Flash, delivering substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks

Shared on Bluesky by 2 AI experts

Originally reported by huggingface.co

Read the original article →

Original headline: DeepSeek Ships V4.1-Flash-1017 With Live Web Search API on 552B Agent MoE