DeepSeek’s V4.1 Flash update posts broad benchmark gains; one small test found it faster than V4 Pro

DeepSeek released V4.1-Flash-1017 on October 17 as an update to its V4.1 Flash model, saying it improves reasoning, coding, agent and visual performance while cutting task time and output length. The company says the update is available on its website, app and API, and the weights are published on Hugging Face and ModelScope.

DeepSeek’s model card reports higher scores than the earlier Flash version in all 11 benchmark rows it lists. The largest gains include Terminal-Bench 4.0, which rises from 31.2 to 42.3, and ProgramBench, from 20.3 to 29.1; DeepSWE v1.1 moves from 74.2 to 77.0. DeepSeek says it used Claude Code for Terminal-Bench and the Mini-SWE-V2 harness for DeepSWE, so these are company-reported results rather than independent evaluations.

Comparisons with other models in DeepSeek’s table are mixed. The new Flash scores higher than the listed Opus 5 results on DeepSWE, Automation-Bench, Agents’ Last Exam and Apex Bench, but lower on the other seven listed tests, including HLE, Terminal-Bench and all three visual benchmarks. The visual scores are marked as using tools and should be read as results from a model-and-tools setup.

A separate review by API reseller OmniaKey reported that 1017 matched V4 Pro on the four tasks both models could run, and completed each faster in its five-task test. The tasks covered coding a Python parser, parallel tool calls, invoice extraction, a long service log and chart reading; V4 Pro could not run the image task. OmniaKey calculated that the Flash model cost less on the shared tasks using DeepSeek’s off-peak list prices. But the reviewer explicitly describes the exercise as a spot check: it used one run per task, through OmniaKey’s endpoint, and does not establish general performance across workloads.

The release continues DeepSeek’s emphasis on efficient inference. Its earlier V4.1-Flash announcement described a 552-billion-parameter mixture-of-experts model with 8 billion active parameters for input and 16 billion for output, and said its KV cache uses a quarter of the HBM and an eighth of the SSD storage required by the previous generation. Those are claims about the underlying V4.1 Flash architecture; the 1017 model card characterizes this release as a minor update and does not provide a separate architecture breakdown.

The company also points to project-building examples, including a web racing game, a Godot recreation of Portal 2 and 3D scenes. DeepSeek says reinforcement learning with self-feedback supports end-to-end project development. The examples show selected outputs, not a general success rate for building projects.

For developers, DeepSeek says 1017 is optimized for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and recommends Harness v0.2.1-rc.1 or later. The model card specifies a 1-million-token context window, gives local inference instructions and lists the weights under the MIT license. It also includes prompt-encoding reference code and a Rust toolkit with Python bindings; inference, tool execution and HTTP transport remain the caller’s responsibility.

DeepSeek opened its Search and Fetch APIs to all users at $0.75 per 1,000 calls for each endpoint. The company says server-side calls through DeepSeek Harness, Claude Code and Codex are currently free for DeepSeek API users, with billing scheduled to start October 31, 2026, at 16:00 UTC. The launch materials do not establish that faster completion or shorter output will lower costs for every use case; the actual economics depend on the workload and tools used.