If you've been following the LLM race, you know how fast your budget evaporates when generating code through commercial APIs. Against that backdrop, the DeepSeek-V4.1-Flash-1017 update from October 17, 2026 feels like a real gift: open weights under the MIT license are now freely available on Hugging Face.
Its Mixture-of-Experts (MoE) architecture packs 552 billion total parameters, activating just 16 billion per token (and 8 billion during prefill), with a context window reaching 1,000,000 tokens.
The biggest leap shows up in coding benchmarks: its Terminal-Bench 4.0 score jumped from 31.2 to 42.3, and on the engineering-focused DeepSWE v1.1 it climbed to 77.0. On the Artificial Analysis leaderboard, the model took 2nd place for intelligence out of 121 systems while outputting 227 tokens per second.
For deeper reasoning tasks, the API now exposes an explicit thinking mode. Best of all is the economics: starting at $0.1 per 1M input tokens, with up to a 98% discount on prompt cache reads. You can easily spin it up on your own cluster via vLLM or SGLang without locking yourself into someone else's cloud.
