DeepSeek Shipped Its Fourth Model in 11 Weeks. Rivals Haven't.

DeepSeek doesn't do press tours. It drops weights on a Saturday afternoon and lets the benchmark table argue for itself.

On Saturday, Oct. 17, the Hangzhou lab put DeepSeek-V4.1-Flash-1017 on Hugging Face, calling it a dated refresh rather than a new model generation. The weights ship as safetensors files split across 48 files, using FP8 8-bit quantization, tagged for text and image-text pipelines in the Hugging Face transformers library. Translation: you don't wait for an API key. You pull it today and run it on your own hardware.

The headline is DeepSeek's own self-graded score. On Terminal-Bench 4.0, a benchmark measuring agentic command-line task completion, the new release scores 42.3 versus 31.2 for the original V4.1-Flash. ProgramBench goes from 20.3 to 29.1 and Apex Bench from 64.6 to 72.1. The technical report claims "substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks."

Translation: we patched this fast enough that last month's numbers are now an embarrassment.

None of those scores come from a neutral referee. Treat them like a vendor's slide deck until a third party runs the same suite.

The cadence closed labs can't match

Here's what you can't wave away: the artifact is real enough that other people are already taking it apart. An independent developer published a quantized GGUF build, benchmarking its degradation against the FP8 weights. That's not hype signaling. That's a stranger on the internet stress-testing your weights because they're actually downloadable.

DeepSeek-V4-Flash shipped July 31, 2026. DeepSeek-V4.1-Flash followed on Sept. 10, 2026, with a 1 million token context window and 552 billion parameters. Now the 1017 refresh lands five weeks later. Four releases in the V4 line inside eleven weeks.

Nobody at OpenAI or Anthropic is shipping open weights on that schedule.

The honest caveat: DeepSeek is not winning outright. On HLE without tools the Chinese model scores 42.6 and Opus 5 56.3, almost fourteen points more. It stays below Opus 5 on Terminal-Bench 4.0 (42.3 vs 51.8) and most vision tests. It wins against OpenAI's cited comparison model on agentic and automation suites, and it wins on price and access against everyone.

Why the lab, not the claims, is the story

DeepSeek is based in Hangzhou and owned by High-Flyer, a Chinese hedge fund, not chasing per-token API revenue the way OpenAI and Anthropic do. Its API for this release starts at $0.15 per million tokens. That's not a discount. That's a different business model: the weights are the product, and the lab profits from everything built around them instead of gatekeeping the model itself.

You do not need DeepSeek's benchmarks to be exactly right to feel the pressure. A usable, cheap, self-hostable model is shipping on a six-week cadence while the labs charging you per token are still arguing about whether open weights are a business model or a liability.

What to watch: whether an independent leaderboard runs V4.1-Flash-1017 through its own suite in the coming weeks, whether enterprises actually swap production traffic onto it, and whether Meta or any American lab answers with its own open release before DeepSeek ships whatever comes next.