DeepSeek V4.1 Flash Launches With 552B MoE, Native Vision and Lower API Prices
DeepSeek has introduced DeepSeek-V4.1-Flash, the smallest model in what the company describes as a new architecture family, combining native visual understanding, a 552-billion-parameter mixture-of-experts design, significantly smaller KV-cache requirements and lower API pricing.
The launch is notable because DeepSeek is not positioning V4.1-Flash as a simple refresh of an existing model. In its announcement, the company says the new family is designed for greater capability, faster inference, higher throughput and scaling to larger models. V4.1-Flash is the first public example of that direction.
DeepSeek says the model uses a new Causal Encoder-Decoder architecture with roughly 8 billion active parameters for input processing and 16 billion active parameters for output generation, despite a total model size of 552 billion parameters. The company also says V4.1-Flash now supports native multimodal input, is live through the DeepSeek API under the deepseek-flash model name, and replaces the older V4-Flash and V4-Flash-Vision-Exp models.
The other major part of the release is cost. DeepSeek reports that V4.1-Flash requires only 890 bytes of global KV cache per token, down from 3,514 bytes for V4-Flash. It also lists off-peak API pricing as low as $0.003 per million cached input tokens, $0.15 per million uncached input tokens and $0.60 per million output tokens.
Taken together, those changes make V4.1-Flash less about winning one isolated benchmark and more about the economics of running coding agents, research systems, tool-using workflows and other applications that repeatedly process large amounts of context.
Key takeaways
- 552B total parameters: V4.1-Flash is a mixture-of-experts model, with DeepSeek reporting about 8B active parameters for input and 16B for output.
- Native visual understanding: multimodal support is now built into the Flash model instead of requiring a separate vision model identifier.
- Stronger coding and agent results: DeepSeek reports 74.2 on DeepSWE v1.1, a 3,471 Codeforces rating and 90.6 on Terminal-Bench 2.1.
- Much smaller KV cache: the reported 890 bytes per token is about 3.9 times smaller than V4-Flash and dramatically below earlier DeepSeek generations.
- Lower serving prices: off-peak API rates are 50% of peak rates, with cache-hit input priced especially aggressively.
-
API migration: developers can use
deepseek-flash; older V4 Flash identifiers temporarily route to V4.1-Flash for compatibility.
A 552B MoE model with asymmetric active parameters
The first number that attracts attention is 552 billion parameters. In absolute terms, that is not a small model. DeepSeek is calling V4.1-Flash the smallest model in its new architecture family, not a small model by conventional parameter-count standards.
The distinction matters because V4.1-Flash is a mixture-of-experts, or MoE, model. MoE systems contain many parameters but activate only part of the network for a given token or workload. That makes the total parameter count less useful by itself for estimating inference cost.
DeepSeek says V4.1-Flash uses its new Causal Encoder-Decoder architecture and activates just 8 billion parameters for input and 16 billion for output. That asymmetric design is one of the most interesting details in the release.
Prompt processing and token generation are different workloads. A large agent can ingest a long repository, browser state, tool output, prior conversation history or documents before generating a comparatively small response. If an architecture can use a lighter active footprint for the input side while allocating more active capacity to generation, it creates another path toward reducing inference cost without simply shrinking the entire model.
DeepSeek has not disclosed enough in the launch thread to independently reconstruct how every routing or expert-selection mechanism works, so it is important not to infer more than the announcement supports. What the company does explicitly claim is that the design is intended to deliver higher capability, faster inference, higher throughput and better scaling to larger models.
DeepSeek also attributes V4.1-Flash's performance to new pre-training methods and larger-scale reinforcement-learning post-training. That combination of architecture and post-training is reflected most clearly in coding and agent benchmarks, where the company reports large jumps over its previous Flash model.
DeepSeek V4.1 Flash benchmarks
DeepSeek's launch materials compare V4.1-Flash against the previous V4-Pro and V4-Flash models as well as GLM 5.3, Kimi K3, GPT-5.6 Sol and Claude Opus 5. The results are company-reported and should be treated as launch benchmarks rather than independently reproduced evaluations, but they show where DeepSeek believes the new model is strongest.
| Benchmark | V4.1 Flash | V4 Pro 0813 | V4 Flash 0731 | GPT-5.6 Sol | Claude Opus 5 |
|---|---|---|---|---|---|
| GPQA Diamond | 90.9 | 92.4 | 89.9 | 94.1 | 93.4 |
| Codeforces Rating | 3,471 | 3,348 | 3,289 | — | — |
| Terminal-Bench 2.1 | 90.6 | 87.9 | 82.7 | 88.8 | 89.1 |
| Terminal-Bench 3.0 | 30.0 | 11.8 | 7.6 | 34.4 | 43.3 |
| Terminal-Bench 4.0 | 31.2 | 12.4 | 7.0 | 39.9 | 51.8 |
| DeepSWE v1.1 | 74.2 | 62.7 | 54.4 | 73.0 | 74.0 |
| ProgramBench | 20.3 | 15.5 | — | 23.0 | 37.0 |
The strongest generational gain appears on DeepSWE v1.1. V4.1-Flash reaches 74.2 in DeepSeek's table, up from 54.4 for V4-Flash and 62.7 for V4 Pro. That is a 19.8-point increase over the earlier Flash model. DeepSeek's comparison also places the result just above the 73.0 listed for GPT-5.6 Sol and the 74.0 listed for Claude Opus 5.
Terminal-Bench paints a more nuanced picture. V4.1-Flash scores 90.6 on Terminal-Bench 2.1, which is above the comparison values DeepSeek lists for V4 Pro, V4 Flash, GPT-5.6 Sol and Claude Opus 5. On the more difficult Terminal-Bench 3.0 and 4.0 versions, however, the model reaches 30.0 and 31.2. Those results are huge jumps over the older DeepSeek models, but the table shows GPT-5.6 Sol at 34.4 and 39.9 and Claude Opus 5 at 43.3 and 51.8.
That is why it would be misleading to reduce the launch to a claim that V4.1-Flash is simply “the best” model. DeepSeek's own data shows a model that becomes far more competitive on software engineering and agent tasks while still trailing other frontier systems on several harder evaluations.
The pattern is similar on GPQA Diamond. V4.1-Flash scores 90.9, slightly above the previous Flash result of 89.9 but below V4 Pro at 92.4, GPT-5.6 Sol at 94.1 and Claude Opus 5 at 93.4 in DeepSeek's table. The model therefore appears to prioritize efficiency and strong practical agent performance rather than maximum reasoning scores at any cost.
On Codeforces, DeepSeek reports a rating of 3,471, up from 3,348 for V4 Pro and 3,289 for the previous Flash model. The launch graphics also show V4.1-Flash at 54.8 on Automation-Bench, ahead of comparison values shown for Kimi K3, GLM 5.3, GPT-5.6 Sol and Claude Opus 5.
The important takeaway is not that one benchmark determines the winner. It is that DeepSeek appears to have pushed its Flash tier closer to flagship-level software-engineering capability while simultaneously reducing the memory footprint and API cost required to run it.
The 890-byte KV cache may be the biggest technical story
For many production teams, the most consequential number in the announcement may not be a benchmark score at all. It may be 890 bytes per token.
During transformer inference, key-value cache data stores information from previously processed tokens so the model does not need to recompute the entire context every time it generates another token. That is essential for efficient generation, but it can consume large amounts of memory as context windows, concurrency and persistent sessions grow.
DeepSeek's launch chart shows a dramatic progression in global KV-cache requirements:
| Model | Global KV cache per token |
|---|---|
| DeepSeek V1 | 389,120 bytes |
| DeepSeek V3.2 | 48,068 bytes |
| DeepSeek V4-Flash | 3,514 bytes |
| DeepSeek V4.1-Flash | 890 bytes |
Using DeepSeek's numbers, V4.1-Flash's cache footprint is approximately 3.95 times smaller than V4-Flash. Compared with V3.2, it is about 54 times smaller. Compared with DeepSeek V1, the difference is roughly 437 times.
DeepSeek summarizes the generation-over-generation infrastructure change another way: compared with the previous generation, the V4.1-Flash KV cache needs only one-quarter of the HBM and one-eighth of the SSD storage.
Those ratios should not be translated directly into claims that every workload becomes four or eight times cheaper. Real production cost depends on hardware utilization, batching, concurrency, context length, cache-hit rate, storage architecture and many other variables. But it is still a meaningful change for workloads where cached context is a major cost center.
This is especially relevant for autonomous or semi-autonomous agents. A coding agent may repeatedly revisit repository context. A research agent can carry long documents and tool outputs across many steps. A support agent may retain a lengthy customer history. In those systems, the model may process far more historical context than it generates as new output.
Reducing cache memory per token can therefore improve the economics of long-lived sessions in a way that does not always show up in a standard benchmark chart.
DeepSeek V4.1 Flash API pricing
DeepSeek pairs the cache reduction with lower absolute API prices. According to the pricing graphic in the launch thread, the new rates took effect at 04:00 UTC on September 10, 2026.
| Token category | Off-peak | Peak hours |
|---|---|---|
| Input — cache hit | $0.003 | $0.006 |
| Input — cache miss | $0.15 | $0.30 |
| Output | $0.60 | $1.20 |
The prices shown are per one million tokens. DeepSeek says off-peak rates are 50% of peak rates. Peak periods are listed as 01:00–04:00 UTC and 06:00–10:00 UTC on weekdays. All other times are off-peak, including weekends and public holidays.
The pricing structure makes cached input unusually inexpensive relative to generation. At the off-peak rate, one billion cache-hit input tokens would cost $3 based on the posted rate, while one billion output tokens would cost $600. That enormous difference reinforces the value of systems that can reuse context effectively instead of continually paying uncached-input or output costs.
As a simple hypothetical example, consider an agent workload that processes 500 million cache-hit tokens, 5 million cache-miss input tokens and 5 million output tokens during off-peak hours. Using the announced rates, the total would be about $5.25: $1.50 for cache hits, $0.75 for cache misses and $3.00 for output. At the corresponding peak rates, the same token mix would be about $10.50.
That example is not a prediction of what a real application will cost, but it illustrates why DeepSeek is emphasizing cache efficiency. For high-context workloads, the cost structure can be dominated by whether context is reused efficiently and how much output the system generates.
Native multimodal support arrives in the Flash model
V4.1-Flash is also now live with native multimodal support. DeepSeek's announcement says developers should set the model to:
deepseek-flash
The older V4-Flash and V4-Flash-Vision-Exp models are retired. For compatibility, DeepSeek says the older deepseek-v4-flash and deepseek-v4-flash-vision-exp identifiers will temporarily route to V4.1-Flash.
That should make migration less disruptive for existing applications, but developers should still update production integrations to the current model name rather than depending indefinitely on temporary routing behavior.
The native visual-understanding capability is strategically important because many useful agent tasks are not purely text based. A model that can reason over screenshots, charts, documents, interfaces and visual state can participate in broader automation workflows without requiring developers to switch to a separate experimental vision model.
DeepSeek's launch material shown here does not spell out every production limit for visual input, so teams should verify current API documentation for details such as accepted image formats, request limits and any modality-specific constraints before deploying multimodal workloads at scale.
Open-source inference support and large-scale deployment
DeepSeek is also pushing V4.1-Flash beyond its hosted API. The company says it will work closely with the open-source community on V4.1-Flash inference support and explore additional deployment options.
A Hugging Face repository is available under deepseek-ai/DeepSeek-V4.1-Flash. DeepSeek's launch post also directly addresses organizations considering very large deployments, asking teams planning infrastructure with 2,000 GPUs plus a storage cluster to get in touch.
That deployment language is significant because it suggests DeepSeek expects V4.1-Flash to be relevant not only as a low-cost public API but also for organizations that want large private or self-managed inference installations.
There is an important wording distinction, however. The launch material says DeepSeek is “supporting open source” and points to the Hugging Face repository. The screenshots themselves do not establish every license term. Organizations considering redistribution, modification or commercial self-hosting should check the repository's current model card and license before making assumptions about what rights are granted.
Why V4.1 Flash matters for AI agents
The broader significance of V4.1-Flash becomes clearer when the release is viewed through the economics of agents rather than chatbots.
A traditional chatbot interaction might contain one prompt and one answer. A modern agent can perform dozens or hundreds of steps: reading files, calling tools, inspecting browser state, executing code, checking results, revisiting earlier context and deciding what to do next. That kind of workflow can consume a huge amount of context even if each individual response is short.
DeepSeek's launch appears designed around that reality. The company is simultaneously reducing active compute, shrinking KV-cache requirements, lowering cache-hit pricing and improving terminal, coding and automation benchmarks. Those are different technical levers, but they all point toward the same product goal: making capable agents cheaper to run continuously.
This is also why benchmark leadership alone does not fully describe the model's value. DeepSeek's own comparison table shows GPT-5.6 Sol and Claude Opus 5 ahead on several difficult reasoning or terminal evaluations. V4.1-Flash does not erase that gap everywhere. Instead, its argument is that developers may not need the absolute highest score on every benchmark if they can get strong enough capability at dramatically lower serving cost.
For high-volume applications, that tradeoff can be decisive. A model that is slightly weaker on one evaluation but several times more economical in a workload's dominant cost center may be the better production choice. Conversely, workloads that require the strongest possible reasoning or reliability on the hardest tasks may still justify using a more expensive frontier model.
The right comparison is therefore not simply “which model has the highest score?” It is “which model delivers the best combination of capability, latency, reliability and cost for this workload?” V4.1-Flash is a more serious competitor because DeepSeek is attacking that entire equation at once.
The bottom line
DeepSeek-V4.1-Flash is an ambitious efficiency-focused release. It combines a 552B MoE model with an unusual asymmetric active-parameter design, native visual understanding, substantially stronger coding and agent benchmarks, an 890-byte KV cache per token and aggressive API pricing.
The reported benchmark results show meaningful progress without making V4.1-Flash a universal leader. It beats or closely matches some flagship models on selected coding and agent evaluations while still trailing them on several harder benchmarks. That balance is important, because it frames V4.1-Flash as a model optimized for practical capability per dollar rather than a model whose only goal is topping every leaderboard.
The cache numbers may ultimately matter most. If DeepSeek's architecture lets organizations sustain long-context agents with much lower memory and storage overhead, the release could reduce one of the less visible costs of persistent AI systems. Pair that with cheap cache-hit tokens and native multimodal input, and the model becomes particularly interesting for coding assistants, research agents, support automation and other long-running workflows.
DeepSeek also calls V4.1-Flash the smallest model in a new architecture family, which implies that the architecture may extend beyond this release. The company has not provided complete specifications or a release schedule for larger members in the materials shown here, so anything beyond that remains speculative.
For now, V4.1-Flash is a strong signal of where DeepSeek is heading: toward models that try to make advanced capability cheaper not only through API pricing, but through the architecture and memory system underneath the API itself.
Source note: This article is based on DeepSeek's V4.1-Flash launch thread and accompanying benchmark, KV-cache, API and pricing graphics posted September 10, 2026. Benchmark figures are company-reported launch results and may differ from independent evaluations or results produced with different harnesses and settings.
View DeepSeek's launch thread · DeepSeek-V4.1-Flash on Hugging Face