Editorial illustration for DeepSeek releases V4.1-Flash-1017, scores 77.0 on DeepSWE, opens Search API
DeepSeek V4.1-Flash-1017 Scores 77.0, Launches Search API
DeepSeek pushed out a new checkpoint of its V4.1-Flash model on October 17, calling it DeepSeek-V4.1-Flash-1017, and put it live on the company's website, app, and API the same day. It's served under the deepseek-flash model name, and DeepSeek says it beats the September V4.1-Flash release on all 11 benchmarks the company reported, with gains ranging from 2.0 to 11.1 points. The top result is a 77.0 score on DeepSWE v1.1, which DeepSeek's own table puts ahead of Claude Opus 5's 74.0. The weak spot is HLE without tools, where the new checkpoint lands at 42.6 against Opus 5's 56.3.
Underneath the benchmark numbers is an architecture built around a 552B backbone paired with 196B Engram parameters, activating 8B parameters per token during prefill and 16B during decode. Context runs to 1M tokens, with output capped at 384K. DeepSeek is pitching the release as a straight upgrade on cost and speed rather than a tradeoff, and paired it with something separate: opening its Search and Fetch APIs to any developer, not just internal use.
DeepSeek has just released DeepSeek-V4.1-Flash-1017, a new checkpoint of its V4.1-Flash model. It is live on the DeepSeek website, app and API. DeepSeek claims stronger scores, faster task completion and shorter outputs than the September V4.1-Flash release.
Why this matters
DeepSeek's numbers are worth taking at face value and no further. A 77.0 on DeepSWE v1.1 beating Claude Opus 5's 74.0 looks good on a chart DeepSeek built itself, but the same model drops to 42.6 on HLE without tools against Opus 5's 56.3. That's a 13.7-point gap in the other direction, and it tells us this is a model tuned for agentic coding tasks, not general reasoning.
For builders, the real story is the Search and Fetch API going public. Pairing a cheaper, faster coding checkpoint with a first-party retrieval layer under one API lowers the cost of standing up an agent that can both write code and pull live information, without stitching together three vendors. That's the part we'd actually test before trusting the benchmark table.
Watch whether independent evals on HLE-style tasks hold up once developers outside DeepSeek start running V4.1-Flash-1017 against their own workloads, and whether the open Search API pricing stays cheap once usage scales past the launch promotion.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives Mathematicians Criticize OpenAI's Navier-Stokes Proof Release
From Archives StepFun Releases Step 5 Preview Open Weights: 600B MoE, 27B Active, 1M Context
Common Questions Answered
How does DeepSeek-V4.1-Flash-1017 perform compared to the September V4.1-Flash release?
DeepSeek-V4.1-Flash-1017 beats the September V4.1-Flash release on all 11 benchmarks reported by the company, with performance gains ranging from 2.0 to 11.1 points. The most notable improvement is a 77.0 score on DeepSWE v1.1, which surpasses Claude Opus 5's 74.0 on the same benchmark.
What is the significance of DeepSeek's 77.0 score on DeepSWE v1.1?
The 77.0 score on DeepSWE v1.1 represents DeepSeek-V4.1-Flash-1017's strongest benchmark result and positions it ahead of Claude Opus 5's 74.0 on that specific test. However, this score should be contextualized as the model is specifically tuned for agentic coding tasks rather than general reasoning capabilities.
What new capabilities did DeepSeek introduce with the V4.1-Flash-1017 release?
Beyond performance improvements, DeepSeek opened its Search and Fetch API to the public as part of the V4.1-Flash-1017 release. This API allows developers to integrate search functionality with the faster and more cost-effective Flash model for building applications.
Why does DeepSeek-V4.1-Flash-1017 show different performance on HLE compared to DeepSWE?
DeepSeek-V4.1-Flash-1017 scores 42.6 on HLE without tools compared to Claude Opus 5's 56.3, a 13.7-point gap in the opposite direction from DeepSWE results. This significant variance indicates the model has been specifically optimized for agentic coding tasks rather than general reasoning, making its performance benchmark-dependent.
Further Reading
- Papers with Code - Latest NLP Research - Papers with Code
- Hugging Face Daily Papers - Hugging Face
- ArXiv CS.CL (Computation and Language) - ArXiv
