DeepSeek V4.1-Flash Retires V4 Pro and Collapses Its Model Tiers
DeepSeek has folded its flagship and budget model lines into a single release. The Hangzhou lab published DeepSeek V4.1-Flash on September 10, 2026, an MIT-licensed open-weight model it describes as the smallest member of a new architecture family, and confirmed that traffic to its larger V4 Pro endpoint will be redirected to the smaller model within days.
The handover starts at 04:00 UTC on September 14, 2026. From that point, API calls addressed to deepseek-v4-pro will be answered by V4.1-Flash and billed at V4.1-Flash rates, an arrangement DeepSeek says will hold until a V4.1-Pro model ships. The earlier V4-Flash and V4-Flash-Vision-Exp models are already retired. Their endpoint names survive as temporary aliases that resolve to the new model, and DeepSeek recommends deepseek-flash as the model string for new integrations.
Memory Is the Headline
The central engineering claim is a collapse in the cost of holding context. DeepSeek reports a global KV cache footprint of 890 bytes per token, about a quarter of what V4-Flash required and roughly 437 times smaller than the V1 generation. FP4 KV caching and cross-layer attention reuse do the work, alongside a 1M-token context window and native image-and-text input.
Working backwards from DeepSeek's ratios, V4-Flash sat near 3,560 bytes per token and the V1 generation near 389,000. Each step in the sequence has cut the figure by a factor of four or better, and the reductions compound on agent workloads that re-read long contexts at every step.
A 1M-token window paired with a small per-token cache changes which workloads are practical. Whole-repository code review, long transcript analysis and multi-hour agent sessions all depend on keeping a large working context resident, and the cache figure decides whether that stays affordable at scale.
Cache size also sets concurrency. Memory held per token determines how many sequences a single accelerator can serve at once, so a fourfold reduction in cache footprint translates into headroom for more simultaneous users on the same hardware, whether that hardware sits with DeepSeek or with a customer running the weights in-house.
Cache-hit charges account for a large share of agent operating costs, as DeepSeek notes, because tool-using workflows repeatedly read the same context. Cache-hit pricing starts at $0.003 per million tokens off-peak, which puts the cost of re-reading a fully cached 1M-token context below a cent. That arithmetic, more than any benchmark row, changes what a long-running agent costs to operate.
Standard rates are $0.15 per million input tokens off-peak and $0.60 per million output tokens. DeepSeek lists the backbone at 552B parameters with 8B activated per token during prefill and 16B during decode, so the throughput gain comes from sparsity as much as from the architecture change.
Where DeepSeek V4.1-Flash Lands on Benchmarks
DeepSeek's published scores place the small model ahead of its own flagship across reasoning, coding and terminal work:
| Benchmark | DeepSeek V4.1-Flash |
|---|---|
| GPQA-Diamond | 90.9 |
| Humanity's Last Exam | 36.8 (39.1 on text-only subset) |
| Codeforces rating | 3471 |
| MathArena Apex | 65.6 |
| Terminal-Bench 2.1 | 90.6 |
| DeepSWE v1.1 | 74.2 |
The DeepSWE v1.1 result travels furthest outside DeepSeek's own reporting. The model card puts V4.1-Flash at 74.2 on that agentic coding test against 73.0 for OpenAI's GPT-5.6 Sol, a comparison DeepSeek now draws itself instead of leaving it to third parties. The margin is narrow and vendor benchmark conditions differ, so the number reads as a parity claim more than a clear lead.
DeepSeek attributes the overall performance lead to tests run by multiple parties without publishing the harnesses behind those runs. Independent reproduction of the agentic scores is the open question that the published table cannot settle.
A Codeforces rating of 3471, Terminal-Bench 2.1 at 90.6 and GPQA-Diamond at 90.9 point the same direction. The model is tuned for multi-step execution and tool use rather than single-turn conversation, which is where agent builders spend their budget.
Open Weights, Three Tiers, One Endpoint
Weights for DeepSeek V4.1-Flash are on Hugging Face under the MIT licence, split across 48 shards, with deployment paths documented for vLLM, SGLang and Transformers. The public API exposes low, high and max reasoning tiers so callers can trade latency against deliberation per request. The release is live across DeepSeek's web app, mobile app and first-party API.
Three reasoning tiers push cost engineering down to the individual request. A classification step and a multi-file refactor no longer have to share one model configuration, and a team can cap spending by routing routine calls to the low tier while reserving the max tier for tasks that fail without it.
The MIT terms are the part enterprise buyers will weigh hardest. Permissive licensing removes the usage restrictions attached to most frontier APIs and allows deployment inside a private network, which matters for regulated industries that cannot send prompts to a third party. The counterweight is hardware. A 552B-parameter mixture-of-experts backbone with a 1M-token context sets a demanding floor for self-hosting, so the licence pays off most for organisations that already own GPU capacity.
The Flagship Tier Is Expiring Faster Than Procurement Cycles
V4 Pro did not have a long life. DeepSeek previewed the V4 series on April 24, 2026, shipping MIT-licensed weights for both Pro and Flash alongside API access. Less than five months later, the Pro endpoint routes to a cheaper model. Any team that built a cost model around V4 Pro pricing, or wrote evaluations against its behaviour, has until September 14 before the model underneath a stable endpoint name changes.
That is the practical sting. Endpoint identifiers normally signal stability; here deepseek-v4-pro keeps its name while its behaviour, price and latency profile are replaced. Teams that pinned the alias instead of the version will see output distributions shift without any deployment on their side. Teams that built fallback logic around a provider's flagship tier will find that tier missing.
Migration itself is mostly administrative. Existing code pointing at deepseek-v4-pro or the older Flash aliases keeps running without edits, and billing drops to the Flash rate automatically on September 14. The work sits in validation: re-running regression suites, checking token accounting against the new cache behaviour, and confirming that prompts tuned for V4 Pro still produce usable output on the smaller model.
Annual procurement and architecture reviews assume model generations last long enough to be budgeted. A flagship that goes from preview to retirement in under five months breaks that assumption, and it pushes evaluation harnesses from a yearly exercise into a continuous one. The cost of re-running agent test suites on every routing change is a recurring expense that procurement templates rarely capture.
The competitive read is direct. DeepSeek is stating that its best-performing model is also its cheapest, which undercuts the usual vendor ladder where capability is rationed by price. Rivals that sell premium tiers at a multiple of their fast tiers have to justify the gap. DeepSeek's own hedge is the promised V4.1-Pro, which suggests it still expects to sell a top tier once it has one worth selling.
Why this matters
DeepSeek has put the strongest model it has built behind an MIT licence and a Flash-tier price, which shifts leverage from model vendors toward the teams that deploy them. For buyers, that means a cheaper baseline and a harder planning problem, because the endpoint names they pin no longer promise the behaviour they were evaluated against. The wider signal is tempo. When a flagship can be retired within five months of its preview, model selection stops being an annual decision and becomes an operational one, with evaluations and cost models that have to run as continuously as the deployments they support.
Sources
Change Log | DeepSeek API Docs
DeepSeek | Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient.
AI-generated image.
Related Articles
- DeepSeek V4 Launch Introduces Trillion-Parameter Pro and High-Speed Flash Models
- DeepSeek-V4-Flash-0731 beats V4-Pro on all nine agent benchmarks
- DeepSeek V4 Pro Adoption Surges as US Firms Pivot to Low-Cost AI Models
✔Human Verified
Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.