DeepSeek V4
DeepSeek’s current V4 generation — V4.1-Flash (deepseek-flash) is the live multimodal Flash SKU; V4-Pro-0813 remains on deepseek-v4-pro after 2026-09-14.
DeepSeek’s V4 generation — live Flash is V4.1-Flash-1017 (API id deepseek-flash, upgraded 2026-10-17): Causal Encoder–Decoder MoE, 552B backbone (8B active prefill / 16B decode), native image+text, and 1M context. API id and prices are unchanged from the 2026-09-10 Flash list. MIT weights on Hugging Face. Legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp aliases temporarily route here. deepseek-v4-pro still serves V4-Pro-0813 after 2026-09-14 at unchanged Pro rates (DeepSeek withdrew the planned Flash reroute).
Why DeepSeek V4 matters
Pin deepseek-flash for new DeepSeek API work. V4.1-Flash (2026-09-10) is a new CED MoE with native image+text, MIT weights, and lower Flash list prices. Aliases deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route there. After 2026-09-14, deepseek-v4-pro still serves Pro-0813 at unchanged Pro rates (DeepSeek withdrew the planned Flash reroute). Keep V3/R1 as open-checkpoint and reasoning baselines, not as assumed live API IDs.
Vision · Tool calling · Thinking · Coding
Last reviewed: 23 October 2026
When to choose DeepSeek V4
Decision guidance for architects—not a feature list.
Best for
- Cost-efficient agentic coding
- 1M-context document and repo work
- Native multimodal agent loops at Flash rates
- OpenAI-compatible / Codex-style backends
- Open-weight self-host where license allows
Avoid if
- You require US-only data residency or a published V4.1-Pro SKU (neither is offered on this alias)
- You require US-only enterprise packaging and SLAs
Strengths
DataAIHub qualitative scores from 1 to 5, for architects. Not official benchmark results or a public ranking.
- Agentic coding / value★★★★★
- 1M context★★★★★
- Open weights + API★★★★★
- Multimodal★★★★☆
- Enterprise packaging★★☆☆☆
Ecosystem
Built by
- DeepSeek
V4.1-Flash is DeepSeek’s current hosted Flash generation (native multimodal, 1M context); Pro-0813 remains on deepseek-v4-pro after 2026-09-14 at unchanged Pro rates.
Competes with
Works with
Recommended for
- DeepSeek Models
How V4 API IDs relate to V3/R1 open checkpoints and routing.
Often paired with
- DeepSeek V3
V3 remains a widely deployed open checkpoint; prefer V4 IDs for new API work.
- DeepSeek R1
R1 for dedicated hard reasoning; V4 for general/agent API traffic.
How DeepSeek V4 evolved
Key moments in chronological order.
- Open source
DeepInfer offline inference for V4.1-Flash
DeepInfer (Apache-2.0) provides offline inference for DeepSeek-V4.1-Flash on NVIDIA SM100 and Huawei Ascend 950. The default checkpoint is DeepSeek-V4.1-Flash-1017. The release does not include a KV cache, prefill/decode disaggregation, or replica load balancing.
- Model
deepseek-flash now serves DeepSeek-V4.1-Flash-1017 on the API, web, and app. Calling method and pricing are unchanged. Legacy Flash aliases still route there.
- API
V4 Pro API continues after planned Flash cutover
DeepSeek keeps serving V4-Pro-0813 on deepseek-v4-pro after 2026-09-14 with billing unchanged, reversing the 2026-09-10 plan to route that alias to V4.1-Flash at Flash rates until V4.1-Pro.
- Model
New CED MoE (552B backbone; 8B prefill / 16B decode) with native image+text and MIT weights. API id deepseek-flash. Aliases deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route here. The launch post also planned to route deepseek-v4-pro to Flash from 2026-09-14; DeepSeek later withdrew that cutover.
- Model
V4-Flash-Vision-Exp multimodal API
Experimental deepseek-v4-flash-vision-exp adds image+text input at Flash rates (up to 384 tokens/image). Matches Flash on text; multimodal agents improve vs text Flash. API-only; no open vision weights yet. Files API for free image reuse shipped the same day. Retired 2026-09-10: the alias now routes to V4.1-Flash.
- API
V4 peak/off-peak API pricing live
Peak/off-peak rates replace the prior flat V4 API prices from 16:00 UTC. Peak hours 01:00–04:00 and 06:00–10:00 UTC; off-peak is half of peak.
- Model
V4-Pro-0813 generally available
Pro leaves preview: agent upgrades, thinking effort low/high/max, native Responses API; peak/off-peak prices from 2026-08-16 16:00 UTC.
- Model
Official Flash release with stronger agentic post-training; MIT weights on Hugging Face; same deepseek-v4-flash API ID.
- Deprecation
Legacy chat/reasoner aliases retire
deepseek-chat and deepseek-reasoner stop after Jul 24 2026 UTC; traffic moves to V4 Flash.
- Model
DeepSeek V4 Preview
V4-Pro (1.6T/49B) and V4-Flash (284B/13B) ship with 1M context and thinking modes.
- API
deepseek-v4-pro / deepseek-v4-flash
Official API IDs replace assuming historic chat/reasoner names map to V3/R1 weights.
Overview
Capabilities
- Vision: Yes
- Audio input: No
- Tool calling: Yes
- Thinking: Yes
- MCP: No
- Coding: Yes
- Structured output: Yes
Technical specifications
- Provider
- DeepSeek
- License
- MIT (V4.1-Flash weights; verify other checkpoints)
- Context window
- 1M
- Parameters
- V4.1-Flash MoE 552B backbone / 8B prefill / 16B decode (+196B Engram); Pro-0813 MoE 1.6T/49B on deepseek-v4-pro
- Architecture
- Causal Encoder–Decoder MoE (CSA2); V4-Pro-0813 remains DSA-class on deepseek-v4-pro
- Release
- 2026-09
- Modalities
- Text, Image
- Vision
- Yes
- Audio input
- No
- Tool calling
- Yes
- Thinking
- Yes
- MCP
- No
- Open weights
- Yes
- API
- Yes
- Pricing (input)
- Flash peak cache-miss $0.30 / cache-hit $0.006 per 1M (off-peak ½); Pro-0813 peak cache-miss $1.32
- Pricing (output)
- Flash peak $1.20 per 1M (off-peak ½); Pro-0813 peak $3.96
Supported modalities
Text · Image
Context window
1M (1,000,000 tokens)
Pricing
Flash list prices from 04:00 UTC 2026-09-10. Peak hours 01:00–04:00 and 06:00–10:00 UTC weekdays; off-peak is half of peak. After 2026-09-14, deepseek-v4-pro still serves Pro-0813 at unchanged Pro rates. Verify live rates.
Availability
Pin deepseek-flash. As of 2026-10-17 that id serves DeepSeek-V4.1-Flash-1017 on the API, web, and app. Calling method and pricing are unchanged. Aliases deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route there. deepseek-v4-pro still serves Pro-0813 after 2026-09-14 (DeepSeek withdrew the planned Flash reroute). MIT weights: deepseek-ai/DeepSeek-V4.1-Flash. DeepSeek also opened Search and Fetch APIs on 2026-10-17 at $0.75 per 1,000 successful requests each. Serving engines may lag the weight drop — verify vLLM/SGLang before self-host.
Use cases
- Cost-efficient agentic coding
- 1M-context document and repo work
- Native multimodal agent loops (screenshots / UI)
- OpenAI-compatible / Codex-style agent backends
- Self-hosted open-weight deployments (V4.1-Flash MIT weights)
Strengths
- Native multimodal Flash with 1M context as the default hosted SKU
- Much smaller KV cache vs V4-Flash (official: ~1/4 HBM, ~1/8 SSD)
- Open MIT weights plus lower Flash API list prices from 2026-09-10
- Legacy Flash / Vision-Exp aliases keep old clients working during the cutover
Limitations
- V4.1-Pro is unpublished; Pro-0813 on deepseek-v4-pro continues until DeepSeek gives further notice
- vLLM 0.30.0 (2026-09-22) adds DeepSeek-V4.1-Flash; confirm SGLang and other engines on your version before self-host
- Enterprise packaging thinner than OpenAI/Anthropic
Related guides
Related benchmarks
Related research
- DeepSeek API updates (V4.1-Flash-1017 and Search API)
Original Paper
- DeepSeek-V4.1-Flash
Original Paper
- DeepSeek-V4.1-Flash API notes
Original Paper
- DeepSeek-V4.1-Flash (Hugging Face)
Original Paper
- DeepSeek-V4-Pro GA Release
Original Paper
- Top Open-Source & Open-Weight AI Models 2026
Evaluation
Related GitHub
Related tools
Related rankings
Companies
Explore more models
- DeepSeek R1DeepSeekDeepSeek’s reasoning-focused model trained with reinforcement learning for multi-step math, science, and coding problem solving.
- DeepSeek V3DeepSeekDeepSeek’s MoE general model — strong open-weight performance on coding and knowledge tasks with competitive API pricing.
- Kimi K3Moonshot AIMoonshot’s Kimi K3 family — 2.8T MoE (104B active), 1M context, native vision. Hugging Face weights under the Kimi K3 License (moonshotai/Kimi-K3) remain the downloadable checkpoint, including in GitHub Copilot. Kimi K3.1 Pro (2026-10-23) is a post-trained K3 and the current API flagship: kimi-k3.1-pro, plus kimi-k3.1-pro-highspeed, which Moonshot says is 2–3× faster. Both are on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. Moonshot says full K3.1 Pro weights will be released by 2026-11-04. Kimi Enterprise access to K3.1 Pro is off by default. The launch post does not publish token prices.
- MiMo V2.6XiaomiXiaomi’s MiMo-V2.6 family — MIT-licensed native-omnimodal MoE with 1M context. Flagship open checkpoint is MiMo-V2.6-Pro-RL (1.02T total / 42B activated); Flash-RL is 309B / 15B activated. Hosted API ids: mimo-v2.6-pro, mimo-v2.6-flash, mimo-v2.6-pro-ultraspeed. Xiaomi publishes an Artificial Analysis Intelligence Index score of 46 for Pro; that figure is vendor-published and is not a DataAIHub ranking.
- Muse GlimmerMetaMeta Superintelligence Labs’ Muse Glimmer — Apache-2.0 ~30B dense multimodal agent model for on-device and single-GPU local agents. Sibling to closed Muse Spark and Muse Wonder; distinct from Llama 4.
- Qwen3AlibabaAlibaba’s Qwen3 family spanning Qwen3.8-Max (2.4T MoE / 95B active; live cloud alias qwen3.8-max routes to the 0902 snapshot as of 2026-09-05), open Qwen3.8-27B (dense VLM, Apache-2.0), Qwen3.8-Flash-Next (125B / 6B active multimodal MoE + 51B n-gram embeddings)—a Qwen4 architecture preview—and Qwen3.8-Omni-Flash (hosted 2026-09-17: text/image/audio/video in, text out, 1M context; API id qwen3.8-omni-flash). Production Qwen3.8-Flash on QwenCloud adds 1M-default context and built-in tools atop the Flash-Next design.