DeepSeek V4.1 Flash Launches with Native Vision and 552B Backbone
DeepSeek just dropped V4.1 Flash with a new Causal Encoder-Decoder architecture that slashes agent memory costs by 75 percent.
Models · Source: Hacker News
What happened
DeepSeek released the October 17, 2026 snapshot of V4.1 Flash. It is their first sparse mixture-of-experts model built entirely on a new Causal Encoder-Decoder architecture. The model uses a highly asymmetric compute split to maximize efficiency. It activates exactly 8 billion parameters for input processing and 16 billion for output generation. Both draw from a massive 552 billion parameter backbone. This specific split keeps the per token compute cost extremely low relative to the massive size of the model.
Vision capabilities are now completely native to the architecture. DeepSeek trained visual and text embeddings jointly right from the start of pre-training. This is a major shift from their earlier experimental V4 Flash Vision model, which bolted image understanding on after the fact. The model handles a massive one million token context window. It also supports generating up to 131,072 completion tokens in a single shot.
OpenRouter traffic data shows heavy adoption by developers building agentic coding tools. Hermes Agent is already pushing over ten billion tokens through it. Claude Code follows closely with over seven billion tokens. DeepSeek positions this as the cost-efficient tier of their V4.1 family. However, they report that it actually exceeds their V4 Pro model on raw performance, generation speed, and total task completion time.
Key facts
- 8B — Active parameters on input
- 16B — Active parameters on output
- 552B — Total backbone parameters
- $0.10 — Price per one million input tokens
- $0.50 — Price per one million output tokens
- 1,048,576 — Total context window token limit
- 186 — Peak tokens per second throughput on OpenRouter
Why it matters
Agent workflows just got significantly cheaper to run at scale. The new compressed KV caching system reduces cache memory requirements to roughly a quarter of the previous Flash generation. You can now build long-horizon computer-use agents without burning through your entire API budget just to retain context across steps. The asymmetric parameter activation is the real secret weapon here. It keeps your per token compute costs low while still leveraging a massive backbone for complex reasoning tasks.
Native multimodal training fundamentally changes how we build visual agents. Since image understanding is baked into the base pre-training, visual tasks will see fewer latency spikes and much better context blending. Competitors will have to match this deep integration. They can no longer just strap vision encoders onto text models after the fact and call it a day. The market will demand native multimodal performance at these lower price points.
For builders
Massive cost cuts for agentic memory
Compressed KV caching cuts memory usage by 75 percent compared to the last generation. Builders running persistent agents with long contexts will save heavily on cache read costs. Cloud providers lose their premium margins on memory-heavy workloads.
High throughput for terminal tools
The model hits up to 186 tokens per second on OpenRouter providers like Parasail. This blazing speed makes it ideal for terminal and coding agents that need to iterate fast. Developers building local dev tools win big here while slow legacy models lose out.
Cheap native vision integration
Vision is trained jointly from the start rather than added later. You pay just 10 cents per million input tokens for a natively multimodal model. Startups building document analysis or UI automation can easily undercut expensive legacy OCR pipelines.
My take
I love seeing DeepSeek ruthlessly optimize for builders. An asymmetric parameter split is a brilliant engineering choice to keep costs at a dime per million tokens while still leveraging a 552 billion parameter brain. If you are building coding agents right now, ignoring this model is just setting your runway on fire.
Original reporting: Hacker News. This is my rewrite and opinion.