⚡ DeepSeek V4.1 Flash Goes Open Source: A Brand-New Architecture
DeepSeek has just open-sourced V4.1 Flash. It supports vision understanding natively, and it's the smallest model in an entirely new architecture family: 550B total parameters, with 8B active on input and 16B active on output (you read that right, input and output activate different amounts).
The design goal for this new foundation is straightforward: a higher capability ceiling, faster inference, bigger throughput, and room to scale to much larger models.
The companion DeepSeek Harness was updated to v0.1.5 at the same time, and the model was specifically trained against its different run configurations. In the API, this generation of Flash takes over from the previous Pro: V4 Pro will be retired in an orderly fashion, its API requests will be routed to V4.1 Flash, and billing will follow the new model's pricing.
🚀 DeepSeek V4.1 Flash
The headline gains for V4.1 Flash land on agent tasks. Compared with V4 Pro:
- DeepSWE v1.1: 62.7 → 74.2
- Terminal-Bench 3.0: 11.8 → 30.0
- Automation-Bench: 43.2 → 54.8
- CyberGym: 83.3 → 88.1
Among the comparison models DeepSeek published, V4.1 Flash scores highest on three of them, DeepSWE v1.1, CyberGym and Automation-Bench, beating Kimi K3, GLM 5.3, Opus 5 and GPT 5.6-Sol in the same table.
DeepSeek V4.1 Flash versus other models on agent benchmarks
Progress isn't uniform across the board, though. V4.1 Flash still scores below V4 Pro on GPQA Diamond, and it doesn't top every rival on tests such as Terminal-Bench 3.0. Put precisely: it beats the previous-generation Pro on multiple agent benchmarks, and on some of them it posts the best score among the frontier models listed.
Complete benchmark results for DeepSeek V4.1 Flash
🧩 550B total parameters, only 8B active on input
V4.1 Flash uses an MoE structure: 550B parameters in total, 8B activated during the input stage and 16B during the output stage.
The new structure splits reading input and generating output into two asymmetric halves. Input runs with a smaller activation footprint, output with a larger one. The 550B figure describes the model's total parameter count, while 8B and 16B describe how many parameters actually take part in the computation at each stage. They aren't interchangeable numbers.
According to the release notes, this asymmetric encoder-decoder design aims to raise the capability ceiling while cutting inference cost, lifting speed and throughput, and leaving headroom to scale to larger parameter counts. V4.1 Flash is the smallest model in the new family, and it also uses a new pre-training approach plus much larger-scale reinforcement learning in post-training.
💰 Long context just got cheaper
The other big change is caching. Context the model has already processed can keep part of its intermediate computation in a KV cache. When a later request shares the same prefix and hits the cache, those results are reused instead of recomputed every time.
Per DeepSeek's published figures, the KV cache per token drops from 3,514 bytes in the previous V4 Flash to 890 bytes, roughly a quarter of what it was. Measured against the original DeepSeek V1, it has shrunk to about 1/437.
Per-token KV cache size across DeepSeek's model generations
The release gives two storage figures alongside it: compared with the previous generation, cache-related HBM requirements fall to 1/4 and SSD requirements to 1/8. Note that this refers to cache resources, not a claim that the whole model's deployment memory drops to a quarter.
The Harness update touches caching too: when running V4.1 Flash, the new DeepSeek Harness can update the system prompt while keeping the existing KV cache intact.
Those cost changes show up in API pricing. The new rates took effect at 12:00 Beijing time on 10 September 2026, and still use peak/off-peak pricing, with off-peak set at half the peak rate. All figures are in RMB per million tokens.
Off-peak
- Input, cache hit: ¥0.02
- Input, cache miss: ¥1
- Output: ¥4
Peak
- Input, cache hit: ¥0.04
- Input, cache miss: ¥2
- Output: ¥8
Peak hours run 09:00–12:00 and 14:00–18:00 Beijing time on weekdays. Everything else, weekends and holidays included, counts as off-peak.
For the same one million input tokens off-peak, a cache miss costs ¥1 while a cache hit costs two cents, a 50x gap. Total job cost still has to include missed input and all output, so don't treat the cached rate as the price of every token.
A hypothetical bill: a batch of tasks racks up 100 million cache-hit input tokens, 1 million cache-miss input tokens, and 100,000 output tokens. At the rates above, run entirely off-peak that's 2 + 1 + 0.4 = ¥3.4. The same usage run entirely during peak hours comes to ¥6.8.
🛠️ The Harness gets an update too
DeepSeek Harness is the runtime that lets the model read and edit files, run commands, call tools, and keep working through a task. It ships a web interface where you pick a workspace, kick off a task, and watch the model's execution and results.
This v0.1.5 release is tightly coupled to V4.1 Flash's training. The model was specifically trained and optimised for standard mode, programmatic tool calling (PTC) mode, and minimal mode, covering different styles of tool use.
File handling filled in a lot of gaps. The web interface now accepts image, PDF and other file uploads that the model reads on demand through file tools. The right sidebar browses the workspace file tree and previews Markdown, HTML, PDF, common code files and images. Files the model generates can be viewed right in the sidebar, opened in your system's default app, or located in the file manager.
During execution, the parent agent and resumable sub-agents can talk both ways, so you can feed in extra information or steer the task. The main agent picks the model and reasoning effort for its sub-agents, while you can queue messages, edit or delete pending ones, interject, or stop the task outright.
The new version also adds experimental Agent Teams. The main agent can spin up several team members, split, assign and track work through a shared task list, let members message each other, and finally consolidate the results itself. It's off by default, needs an experimental plugin to enable, and burns extra tokens when used.
Plugin entry points are clearer as well. The left sidebar lets plugins register a global entry that links to the central content panel, and the right side can register new tabs displayed via multiple tabs, split panes or full screen. Loading, restoring, memory usage and history navigation in long sessions were all optimised, with a built-in plugin management panel planned for later.
In the community showcases DeepSeek shared, developers have already used V4.1 Flash with Harness to build a papercut-style game called Paper Tiger, tagged as "made in two rounds of prompting". Another demo shows optical rendering through a triangular prism.
Community showcase: Paper Tiger, a papercut-style game made in two rounds of prompting
Community showcase: triangular prism optical rendering
🧪 How to use it
Per this release, V4.1 Flash is already live on the DeepSeek API under the model name:
deepseek-flash
The older V4 Flash and V4 Flash Vision Exp are being retired. For backwards compatibility, the existing deepseek-v4-flash and deepseek-v4-flash-vision-exp names route to V4.1 Flash for now.
After 12:00 Beijing time on 14 September 2026, and until V4.1 Pro launches, requests sent to deepseek-v4-pro will also be routed entirely to V4.1 Flash and billed at V4.1 Flash rates.
Tencent's WorkBuddy and CodeBuddy, along with OpenCode, are official partners and have fully integrated the new model.
To run DeepSeek Harness on any system with a Node.js toolchain installed:
npx @deepseek-ai/dsh web
Once the web interface starts, add your DeepSeek API key in settings, pick a workspace, and you're off. To install from source, clone the repo and follow its setup instructions:
git clone https://github.com/deepseek-ai/deepseek-harness
Feedback on the model and Harness can be submitted with the thumbs up/down buttons in the interface, or the /feedback command.