None
None
Long Story Video Skill
Turn your idea into 10s–10min video—AI writes scripts, prompts, and generates footage automatically.
Ads Video Skill
Generate professional ads and sales videos—AI auto-generates scripts, prompts, and footage.
3D Science Explainer Video Skill
Convert scientific concepts into stunning 3D explain animations
Feedback
😢 Every free AI generation costs us real GPU money. We pay out of pocket to keep this free for you.
💎 Upgrade to Pro and unlock:
- ✓HD Download
- ✓No Watermark
- ✓Unlimited Generations
deepinfer deepseek
deepinfer deepseek: an Apache-2.0 engine for V4.1-Flash on SM100 and Ascend 950. Not a model or API; benchmark caveats flagged.
All Tools
Discover our comprehensive AI-powered animation toolkit
MiniMax H3
MiniMax H3 AI Video Generator
Seedance 2.5
The Future of AI Video Is Here.
Seedance 2.0
The Future of AI Video Is Here.
Veo3.1
Create Stunning Videos with Veo3.1
FLUX 3 Video Generator
Kling 3.0
Next-Gen AI Video Generator
Grok Video Generator
Create Videos from Text or Images with AI
MiniMax H3 video generator
MiniMax H3 AI Video Generator
What the DeepInfer Release Actually Is
An Apache-2.0 inference engine for DeepSeek-V4.1-Flash — not a model, not a hosted API, and not a complete serving stack.
- Engine, Not ModelApache-2.0 code for running V4.1-Flash, out 20 October 2026; the weights are MIT-licensed and versioned 1017.
- Where the Code StopsThree gaps are named outright: KV cache, prefill/decode disaggregation, replica load balancing — all held back with internal infra.
- Hardware and RecipeSM100 and Ascend 950 are listed, charts are NVIDIA-only, and deepseek-recipe ships separately under MIT for protocol conversion.
How to Run DeepInfer
Three steps, plus the serving work the release leaves to you.
Inside the DeepInfer Engine
The engineering choices the release actually contains: kernels, parallelism, speculation, cache policy, scheduling, and determinism.
Six Kernels, One Engine
The attention, GEMM, expert-communication, tile, and token-selection kernels shipped as separate projects; now they run inside one engine.
Wide-EP with EPLB
Attention goes data-parallel, MoE layers expert-parallel; decode spreads experts over 64–128 GPUs and EPLB rebalances hot ones with redundant copies.
DSpark with a Confidence Head
Draft tokens are proposed ahead and a confidence head decides which are worth verifying; DeepSeek reports +4–7% throughput at its production point.
SWA Bounded Replay
Instead of rebuilding the whole sliding-window cache at every layer, only the recent tail is replayed — approximate output, bounded recompute.
Async Scheduling by Default
The CPU prepares the next batch while the GPU runs the current one, closing the idle gap between decode steps — small detail, real throughput value.
Deterministic Modes on NVIDIA
Batch-invariant inference and deterministic sampling let benchmark runs and agent traces be audited, which is what makes published numbers checkable.
DeepInfer DeepSeek — FAQ
Straight answers on the V4.1-Flash release, hardware support, the missing serving pieces, and the benchmark charts.
Is DeepInfer a model or an engine?
An engine. It runs DeepSeek-V4.1-Flash-1017, whose weights are MIT-licensed, while the DeepInfer code itself is Apache-2.0 and allows commercial use with the usual notice obligations.
Can I run DeepInfer on an RTX 4090 or H100?
Not per the documentation. The supported list covers SM100 GPUs and Ascend 950 NPUs; anything older from NVIDIA, or a consumer card, is not on it.
Does DeepInfer ship an OpenAI-compatible API server?
No. What ships is offline inference. deepseek-recipe covers protocol conversion and tokenization but is not a server, so the serving layer remains your build.
Why is the KV cache system missing?
DeepSeek describes the cache-management and load-balancing code as too entangled with its own infrastructure to publish in full, so that portion stays withheld for now.
How fast is DeepInfer, and what do the charts prove?
Two Pareto charts were published — prefill on 4 SM100 GPUs and decode on 64 at a 131,072-token context — with no outside reproduction, so treat them as trade-off maps, not promises for your workload.
Which DeepInfer is this page about?
DeepSeek's 2026 inference framework. The same name also labels an unrelated 2016 medical-imaging toolkit and a 2023 smart-contract type-inference project; DeepInfer is not affiliated with the provider DeepInfra.
Go Deeper Into DeepInfer
Continue to what shipped, how it works, benchmarks, limits, and the name mix-up — independent, never run here, every figure tied to a public source.