DeepSeek ships V4.1-Flash-1017: 77.0 on DeepSWE, MIT weights, and a $0.75 Search API

• Who: DeepSeek — announced October 17, 2026 on X and in its API docs, with weights posted to Hugging Face the same day.
• What: DeepSeek-V4.1-Flash-1017, a minor update to V4.1 Flash delivering improvements across reasoning, coding, agentic, and visual benchmarks.
• Headline score: 77.0 on DeepSWE v1.1 (up from 74.2; ahead of Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0). Terminal-Bench 4.0 jumps to 42.3 from 31.2.
• Price: $0.15 per million input tokens / $0.60 per million output tokens, unchanged; 1,048,576-token context, up to 393,216 completion tokens, API name deepseek-flash.
• License: MIT — weights on Hugging Face and ModelScope; companion paper "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression."
• New: Search and Fetch APIs open to everyone at $0.75 per 1,000 successful requests (5 QPS cap per account); free via DeepSeek Harness, Claude Code, or Codex through end of October 2026.

DeepSeek pushed its newest checkpoint, V4.1-Flash-1017, live across its web chat, app, and API on October 17, 2026 — a "minor update" to V4.1 Flash that the company says cuts task execution time and output length while lifting scores across reasoning, coding, agentic, and visual benchmarks. The weights are MIT-licensed, and DeepSeek used the release to open its in-house Search and Fetch APIs to everyone for the first time.

The headline number: 77.0 on DeepSWE

The model card's benchmark table puts DeepSWE v1.1 at 77.0 — up from 74.2 on the previous Flash checkpoint, and ahead of Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0) on that specific coding benchmark. Terminal-Bench 4.0 jumps to 42.3 from 31.2, a sharp move for agent-terminal work. MarkTechPost's read of the release material reports the new checkpoint beats its predecessor on all 11 benchmarks DeepSeek published, by 2.0 to 11.1 points.

Elsewhere on the card: HLE without tools at 42.6, NL2Repo-Bench at 68.2, Apex Bench at 72.1, BabyVision with tools at 92.5, and Automation-Bench at 60.5. DeepSeek ran Terminal-Bench on the Claude Code harness and DeepSWE on Mini-SWE-V2.

Coding yes, raw reasoning no

The fine print matters. On pure reasoning, the frontier still leads: HLE without tools comes in at 42.6 for Flash-1017 versus 56.3 for Opus 5, and GPT-5.6 Sol edges it too at 44.5. This release wins the agent-coding lane — autonomous software work, terminal tasks, repo-level generation — not the raw-reasoning crown. DeepSeek frames it the same way, calling the update an advance of the "performance–cost Pareto frontier across multiple agent harnesses."

Price unchanged — that's the point

The API name deepseek-flash and its pricing are unchanged: $0.15 per million input tokens and $0.60 per million output tokens, with cache reads at $0.003 per million. The context window stays at 1,048,576 tokens with up to 393,216 completion tokens, and the model card recommends temperature 1.0 and top_p 1.0. Faster tasks and shorter outputs on the same price card is the whole pitch: better Pareto, not a new price.

A live Search API, priced to be used

Alongside the weights, DeepSeek opened its Search and Fetch APIs — backed by its in-house search engine — to everyone at $0.75 per 1,000 successful requests, capped at 5 QPS per account. Server-side calls routed through DeepSeek Harness, Claude Code, or Codex stay free through the end of October 2026, after which the rate card applies. It's the company's first public rate card for live web search inside its own models — a direct play at the agent-search lane.

Under the hood: KV cache compression

Technical coverage of the release describes the checkpoint as the first built on DeepSeek's Causal Encoder-Decoder (CED) architecture — a 552B-parameter sparse mixture-of-experts design with asymmetric activation to keep per-token compute low. The companion paper, titled "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression," puts the memory story front and center: compressing the KV cache is what lets a model this size serve agent workloads at fifteen cents a million tokens. The company also says reinforcement learning with self-feedback improved end-to-end project building, such as game creation and 3D modeling.

Why it matters

Agent coding is where inference spend actually lands — long multi-step trajectories, terminal loops, tool calls — and DeepSeek just made its strongest entry in that lane cheaper per task without touching the price card. MIT weights mean the 552B checkpoint can be served, distilled, and studied anywhere the hardware allows, and the community moved fast: a GGUF conversion sized for a single 128GB DGX Spark appeared on Hugging Face the same day. The open question is how quickly rivals answer the DeepSWE number — and whether anyone answers the Search API.

Sources: the official DeepSeek-V4.1-Flash-1017 model card on Hugging Face (benchmarks, license, sampling parameters), AI Weekly's release coverage (Search API rate card, MarkTechPost benchmark deltas), and OpenRouter's model page (pricing, context window, API name).