DeepSeek-V4.1-Flash Official Launch: deepseek-flash API From $0.15/$0.60 per 1M, 1M Context, MIT Open Weights

What's confirmed (both sources checked Sep 10, UTC+8)

• The release: DeepSeek's change log entry dated 2026-09-10 announces the official release. V4.1-Flash is "the smallest model in our new architecture family, with native multimodal visual understanding"; the architecture targets a higher capability ceiling, faster inference, higher throughput, and scaling to larger models.
• The model card: the Hugging Face README ("Pushing the Limits of KV Cache Compression") specs a 552B-backbone multimodal MoE — 1 shared + 384 routed experts, 6 activated per token — activating 8B parameters/token at prefill and 16B at decode, with a 1M-token context trained on 45T tokens (sparse attention at 64K, extended to 1M at 34T).
• Architecture headline: a Causal Encoder-Decoder (CED) — 20-layer causal encoder + 20-layer decoder — projects the decoder's global KV cache from final encoder states; FP4 main KV caching brings the global cache to ~890 bytes/token, ≈1/4 of DeepSeek-V4-Flash (≈437× less than V1), while SWA Bounded Replay cuts the persistent cache to ≈1/8.
• Headline benchmarks (official, max reasoning effort): GPQA Diamond 90.9, Codeforces rating 3471, Terminal-Bench 2.1 90.6, DeepSWE v1.1 74.2, HLE w/ tools 63.9, Agents' Last Exam 31.8. Reasoning effort is continuously controllable (1–100). Vendor-run numbers — treat as such until third-party replication lands.
• API name and aliases: new requests should use deepseek-flash. deepseek-v4-flash and deepseek-v4-flash-vision-exp are "still accepted" but temporarily routed to V4.1 Flash — the V4-generation models they named have been retired.
• V4 Pro retirement schedule: after 12:00 Beijing time (04:00 UTC) on September 14, 2026, and until the future release of V4.1 Pro, all deepseek-v4-pro requests are routed to V4.1 Flash and billed at the V4.1 Flash price. DeepSeek frames it as an "orderly retirement" after testing showed V4.1 Flash beats V4 Pro on performance, cost, speed and total time.
• Price cuts are official: "With the release of DeepSeek-V4.1-Flash, API prices have been reduced accordingly." The verified card: off-peak $0.15/$0.60 per 1M (input cache-miss / output), $0.003 cache-hit input, peak = 2× during 01:00–04:00 & 06:00–10:00 UTC Mon–Fri (09:00–12:00 & 14:00–18:00 Beijing). Concurrency 2,500 vs V4 Pro's 500. Off-peak discount fans already know the pattern from the weekend off-peak pricing shift — this makes off-peak the default again.
• Open weights, MIT: the HF repo ships weights under MIT, a technical report PDF, a self-contained prompt-encoding reference (no Jinja chat template), and local inference instructions — same open-weights posture as Vision-Exp's ungated MIT drop, now on the mainline multimodal model.

What didn't ship — and what to watch

• No V4.1 Pro yet. As of Sep 10 (UTC+8), the change log names V4.1 Pro only as a future release bounding the V4 Pro routing switch. No date, no benchmarks. Expect a separate event — the new family "scales to larger models" by design.
• Alias names are temporary. If you still call deepseek-v4-flash or deepseek-v4-flash-vision-exp, you're already getting V4.1 Flash at Flash prices — but the docs mark the routing as temporary. Migrate to deepseek-flash now to avoid a future break.
• Sep 14, 04:00 UTC is the V4 Pro cliff. After that moment, deepseek-v4-pro responses come from V4.1 Flash — behavior and benchmark numbers will differ from V4-Pro-0813 even though the model name doesn't change. Pin your evals before Monday noon Beijing time.
• Third-party coverage pending. Per this site's launch policy we cite only first-party official sources for release facts; independent benchmark replication and overseas/X analysis typically land 24–48h after a DeepSeek launch and are not yet reflected here.

Why the KV cache number matters for builders

• Long-context agents get cheaper where it hurts: input. 890 bytes/token of global KV (vs ~3.5KB on V4-Flash) means million-token contexts stop being a memory bomb. With 8B active params at prefill, input-heavy workloads — repo-scale RAG, long agent trajectories, document pipelines — are exactly the profile this model optimizes.
• The rate card rewards cache discipline. Cache-hit input at $0.003 (off-peak) is 50× cheaper than a cache-miss input call. With the DeepSeek API's implicit caching, stable prefixes (system prompts, tool schemas, long documents) become almost free to reuse.
• Concurrency 2,500. Four times V4 Pro's limit of 500 — a signal about throughput headroom on the new architecture. Rate-limit details live in the official Rate Limit & Isolation page.
• Self-host vs API: MIT weights make the hedge real, but a 552B-backbone model (even at 8B/16B active) is multi-node GPU territory. Most builders should consume the API and keep the weights as a fallback — DeepSeek's own DeepSeek-V4-Pro-for-builders calculus shifts again with V4 Pro itself now scheduled to route into Flash.

Access from outside China — and the usual disclaimer

The official endpoint (https://api.deepseek.com) serves OpenAI-format and Anthropic-format bases, with Responses API, tool calls, JSON output and vision on deepseek-flash. International builders should verify payment rails and billing independently — official top-up options remain China-centric for some payment methods.

Independence disclaimer: ChinaModelAPI is an independent information site, not affiliated with DeepSeek. Every fact and price in this article was transcribed from DeepSeek's official change log and official Hugging Face model card on Sep 10, 2026 (UTC+8); This article is not a purchase recommendation. Verify prices on the official rate card before you build.

Primary sources

• DeepSeek API Docs — Change Log, entry dated 2026-09-10: "DeepSeek-V4.1-Flash Release" (API changes, V4 Pro routing, pricing note)
• Hugging Face — deepseek-ai/DeepSeek-V4.1-Flash model card: architecture, 45T-token pre-training, evaluations, MIT license (raw README, checked Sep 10, 2026)
• DeepSeek API Docs — Models & Pricing: deepseek-flash rate card, peak-hour definition, alias and routing footnotes (checked Sep 10, 2026)

FAQ (2026)