October 19, 2026 | AI | By Mahesh
DeepSeek released V4.1-Flash-1017 on October 17, 2026, a checkpoint update to its V4.1-Flash model that is live on the web, in the app and on the API, with MIT-licensed weights on Hugging Face and ModelScope [1][2]. It keeps a 1 million token context window and the same API name, and DeepSeek says it delivers stronger benchmarks, faster task completion and shorter outputs than the September version [2]. Alongside it, DeepSeek opened standalone Search and Fetch APIs at $0.75 per 1,000 requests. Every benchmark figure so far comes from DeepSeek itself.
- V4.1-Flash-1017 is a minor checkpoint update that DeepSeek says is faster and terser, trained with reinforcement learning on end-to-end projects such as games and 3D scenes.
- Weights are MIT licensed and API prices halve off-peak, with Search and Fetch free through server-side tools until billing starts.
- Gains are vendor-reported with no independent replication, and Claude Opus 5 still leads on harder reasoning and terminal tasks.
What is DeepSeek-V4.1-Flash-1017 and what is new
DeepSeek-V4.1-Flash-1017 is an updated checkpoint of V4.1-Flash, which launched in early September. Bloomberg called it a minor update, and the model card uses the same description [1][4]. Requests to the deepseek-flash endpoint and the legacy deepseek-v4-flash names now route to the new version, so existing integrations need no changes [3][5].
DeepSeek says reinforcement learning with self-feedback improved end-to-end project development, and its demos include a 3D model of the Forbidden City, a browser racing game and a Portal 2 recreation in Godot. It recommends DeepSeek Harness version 0.2.1-rc.1 or later, and the model is tuned for Tencent CodeBuddy and WorkBuddy [2]. DeepSeek calls the result a Pareto improvement in performance and cost.
"Stronger benchmark performance, faster task completion, and shorter outputs than the previous version."DeepSeek, release notes, October 17, 2026 [2]
The company also made Search and Fetch generally available as standalone APIs, powered by its in-house search engine, which DeepSeek says handles hundreds of millions of requests a day [2]. The release follows a roughly five-hour DeepSeek outage earlier in the week that some experts linked to limited computing capacity [4].
Under the hood: size, context and formats
DeepSeek does not state a parameter count on the model card. SmartScope puts the model at about 763 billion parameters, calculated from the size of the published safetensors files, and active parameters are not disclosed [5]. No new architecture paper has been published [7]. The context window is 1 million tokens, with up to 384,000 tokens of output, thinking and non-thinking modes, tool calls, JSON output and vision support [3].
Weights are stored as safetensors in BF16, F32, F8_E4M3 and I8, and the model card links quantized builds for llama.cpp, Ollama and LM Studio. It points to vLLM and SGLang for serving and gives no hardware requirements, and the repository has no Jinja chat template, so prompts need the reference encoder in its encoding directory [1][5]. SmartScope estimates that practical local use needs hundreds of gigabytes of memory.
Benchmarks: DeepSeek's table against the September model
The scores below come from DeepSeek's model card. No independent lab had published results on the release when the sources used here were written [1][5][6].
| Benchmark (DeepSeek-run) | V4.1-Flash-1017 | V4.1-Flash |
|---|---|---|
| DeepSWE v1.1 | 77.0 | 74.2 |
| Terminal-Bench 4.0 | 42.3 | 31.2 |
| HLE, no tools | 42.6 | 36.8 |
| Apex Bench v1.1 | 72.1 | 64.6 |
The largest gain is on Terminal-Bench 4.0, up 11.1 points, followed by ProgramBench, which rose 8.8 points [6]. DeepSeek's cost chart shows about 77% on DeepSWE v1.1 at roughly $0.20 per task on the maximum setting. Those are approximate readings of a vendor chart that uses DeepSeek's own harness [5].
The comparison set matters. DeepSeek's table pits the model against Claude Opus 5, Kimi K3 and GLM-5.3 and omits Claude Opus 5.5 and the GPT-6 generation [1][6]. Reflection AI's own table for its new Beam model lists V4.1-Flash at 74.2 on DeepSWE, which matches the September figure above [8].
Pricing, availability and how to access it
The deepseek-flash API costs $0.30 per million input tokens on a cache miss, $0.006 on a cache hit and $1.20 per million output tokens at peak. Off-peak rates are half: $0.15, $0.003 and $0.60. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. The concurrency limit is 2,500 [3].
Search and Fetch cost $0.75 per 1,000 successful requests each. Calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free until billing starts on October 31 at 16:00 UTC, which is 00:00 Beijing time on November 1. DeepSeek notes it may change prices [2][3]. Developers can reach the model through the web app, the mobile app or the API documentation, and download weights from Hugging Face.
Limitations, trade-offs and alternatives
On DeepSeek's own comparison, Claude Opus 5 scores higher on HLE, Terminal-Bench 4.0, ProgramBench, NL2Repo, Chartography and BabyVision, with a lead of about 14 points on HLE. The new model leads on DeepSWE v1.1 at 77.0 against 74.0, and on Automation-Bench and Apex Bench [6]. Hacker News commenters were skeptical, and some reported that cheaper models stall on hard debugging. A common pattern is a frontier model for planning and review with a Flash-class model for bulk work [6].
Open questions remain. The sources do not document rate limits for Search and Fetch, nor whether failed or empty calls are billed, and treating the free period as a launch promotion is prudent [7]. For other open-weight options, see Depth Grid's coverage of Mistral Large 4, whose weights are due by the end of October, and of the Nvidia and Reflection AI talks. Reflection's Beam is the newest US open-weight entrant and costs nothing to download.
Sources
- DeepSeek. "DeepSeek-V4.1-Flash-1017," model card, Hugging Face. huggingface.co
- DeepSeek. Release notes, October 17, 2026. api-docs.deepseek.com
- DeepSeek. "Models and Pricing," API documentation. api-docs.deepseek.com
- Bloomberg via Yahoo Finance. "DeepSeek Pushes Out Updated AI Model It Says Is Faster, Leaner," October 17, 2026. finance.yahoo.com
- SmartScope. "DeepSeek-V4.1-Flash-1017: Changes, Benchmark Comparisons, API Pricing, and Specifications." smartscope.blog
- Developers Digest. "DeepSeek V4.1-Flash-1017: Benchmarks vs Opus 5, Pricing, Switch?" developersdigest.tech
- Tech Insider. "DeepSeek V4.1-Flash-1017 Hits 77.0, Opens Search API." tech-insider.org
- Reflection AI. "Introducing Beam," October 5, 2026. reflection.ai
Read More on Depth Grid
- Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing
- Nvidia in Talks to Deepen Stake in or Buy Reflection AI
- Meta Muse Wonder Launch: 1M Context, $4 API Price, Benchmarks
- Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks
Article by Mahesh | Depth Grid
Discussion (0)
Comments are disabled for this post.