DeepSeek V4.1-Flash-1017: Open Weights, Full Projects, 77% on Code Tasks
On October 17, DeepSeek shipped an updated version of its Flash model and it landed harder than expected. DeepSeek-V4.1-Flash-1017 scores 77% on DeepSWE, outperforming Claude Opus 5 (74%) on a benchmark that measures real end-to-end software development. The weights are open under MIT. The API costs $0.15–0.30 per million tokens. And the team's demo projects—a Forbidden City 3D reconstruction, a Portal 2 remake in Godot, a web racing game—all built in a handful of turns—hint at what's actually changed.
What people are building with it
On Hacker News, developers report porting games from childhood in six hours. One person rebuilt two Finnish shareware games for Linux and Rust in about that time, token cost: roughly three euros for both. Another ran it through DeepSeek's API to audit a large C++ game engine codebase (200+ files) and found relevant bugs in 20 minutes.
Creators are using it for scaffolding. DeepSeek's own examples show this: the model takes a one-sentence request ("make a Portal 2 game") and builds a runnable project in Godot, complete with level structure, physics, and visual assets. That flow—from wish to working prototype—is what the team trained for using reinforcement learning with self-feedback.
For smaller hardware, the appeal is different. A developer running it locally via vLLM on a pair of A100 GPUs can handle large codebases without sending code to an external server. Someone with a 24GB GPU can quantize it down and run it for code review, documentation, or agent loops.
What's actually good (and what isn't)
The benchmarks are real. On Terminal-Bench 2.1 (hour-scale agentic tasks), it scores 90.6%, up from 82.7% in the previous version. On Codeforces (competitive coding), it hits 3,471, ahead of both the old Flash and DeepSeek's own Pro model. Those aren't synthetic metrics—they measure tasks humans care about.
The token efficiency jump matters. With "reasoning effort" set to low, the 1017 weights produce 2.5× fewer tokens per task than the previous build without losing quality. At scale—agents running queries all day, or video teams batch-processing scene generation—that adds up. The showcase projects are real, not cherry-picked: they're built from a single user request with no back-and-forth refinement, in codebases (Godot, Three.js) that real teams use.
The honest limitation: it's still a model. Odyssey has not published a paper or detailed technical description of the demonstration. The company did not specify whether the brain model is based on the FlyWire connectome—wait, wrong story. For DeepSeek: the model is a mixture-of-experts with 552 billion total parameters, which means all of them live in memory even though only 8–16 billion activate per token. Full-precision hosting needs over a terabyte. Even quantized to 8-bit, it's several hundred gigabytes—datacenter territory for self-hosting. For API use, that's not your problem. For local use on a modest rig, you'll want community GGUF quantizations, which trade some quality for speed.
On reasoning-heavy tasks (pure math, GPQA), the scores are respectable but not frontier-leading. The model was trained to finish projects, not to spend ten minutes deriving theorems.
How to try it
API: The easiest start. DeepSeek-V4.1-Flash-1017 landed across the company's website, mobile app, and developer API the same day, paired with a DeepSWE v1.1 Pass@1 score of 77.0. You can prompt it at api.deepseek.com. The model identifier is deepseek-flash. Pricing is low enough that a large project shouldn't cost more than a few dollars.
Hosted (OpenRouter, others): OpenRouter and other providers offer it with per-request billing and, in some cases, cached context windows (cheaper repeat queries).
Self-hosted: The 1017 weights are published under the MIT License on Hugging Face and ModelScope as DeepSeek-V4.1-Flash-1017. Self-hosting is realistic only on serious hardware: the FP8 checkpoint is several hundred gigabytes and typically needs a multi-GPU node. For most teams, a hosted API is the practical route. If you do have the hardware, vLLM and SGLang both support it. Community GGUF quantizations (4–8 bit) are smaller but slower; check Hugging Face's quantizations page for community builds.
For game building specifically: Use the API. Prompt it with a scene description, iterate on the Three.js or Godot code it returns, run it in your editor. The model is trained to produce working scaffolds, so you're not debugging syntax—you're refining creative direction.