DeepSeek's 10 Game-Changing Open-Source Projects: From GRPO to Agent Infrastructure

The author systematically reviews the 10 most valuable open-source projects released by DeepSeek, selected based on open-source influence, audience reach, star counts, and repository quality.

10. DeepSeek-Math

Released in February 2024 with only 3K+ stars, this project's historical weight rivals many billion-parameter models. At a time when the industry believed massive SFT data stacking was the only path, DeepSeek-Math 7B started from DeepSeek-Coder-Base and used a custom unsupervised/weakly-supervised cleaning pipeline to extract a 120B-token high-purity math corpus (DeepSeekMath Corpus) from Common Crawl.

This work proved a key fact: continuous pre-training on high-quality code and math corpora is the cornerstone for activating deep logical reasoning in LLMs .

Additionally, the paper introduced GRPO (Group Relative Policy Optimization) , which became the algorithmic foundation for the later globally influential DeepSeek-R1-Zero and R1. This repository is the essential first lesson for understanding today's LLM reinforcement learning.

9. DeepEP

Released on day two of DeepSeek's open-source week in February 2025, DeepEP is the communication library for DeepSeek's MoE models. In modern MoE architectures with hundreds of sparse experts, a single GPU cannot hold all experts; they must be distributed across thousands of GPUs. Each token must be routed across cards or physical nodes to designated experts via two All-to-All collective communications , the primary bottleneck for MoE training and inference speed.

DeepEP is one of the few industrial-grade communication libraries that fuses low-level protocols with specific hardware topology and open-sources it:

Intra-node NVLink, inter-node RDMA : GPUs within a server use high-speed interconnect; cross-server traffic uses low-latency networking to move tokens to experts and gather results faster.

Communication-computation overlap : While transmitting data, schedule independent computation to reduce GPU idle time waiting for network.

FP8 data distribution support : Transmit data in compact 8-bit floating-point format, reducing data movement volume and easing bandwidth pressure.

Anyone building large-scale MoE production clusters or optimizing distributed training schedules should treat DeepEP as the sole industrial-grade reference standard.

8. 3FS

The finale of the 2025 open-source week, 3FS (just over 10K stars) is the storage backbone enabling DeepSeek's 10K-GPU clusters to sustain massive data throughput. Large-model pre-training and ultra-long-context inference face two fatal I/O scenarios:

Massive small-file concurrent throughput during pre-training : Thousands of rank nodes request text and code data from different domains at extremely high concurrency; any stall causes compute idling.

Lightning-fast checkpoint flush and KV state swap-in/out : Hundreds of GB of weight checkpoints must save in seconds, otherwise fault recovery and training progress costs become unacceptable.

3FS completely discards traditional POSIX legacy baggage and redesigns the storage path for AI workloads:

Compute-storage separation : GPU nodes compute; storage nodes pool large numbers of SSDs into shared storage. Compute nodes access data uniformly without storing all training data locally.

CRAQ strong-consistency replication : Multiple replicas of the same data share read load while a replication protocol guarantees consistency.

RDMA high-speed access : Compute and storage nodes exchange data over high-speed networks, reducing transfer overhead and boosting throughput for training data reads, checkpoint saves, and inference cache access.

The 3FS source code (C++ plus Rust, hundreds of thousands of lines) is daunting; the author advises reading only the README and design docs.

7. FlashMLA

A high-performance MLA decoding operator library optimized specifically for NVIDIA Hopper architecture (H800/H100). DeepSeek-V2 and V3 achieved unprecedented inference cost reduction largely due to MLA (Multi-Head Latent Attention) , which compresses the massive KV cache by over 90%.

However, after MLA compression, the bottleneck in some decoding scenarios shifts from data movement to matrix computation; generic operators are insufficient. FlashMLA addresses this by:

Direct adaptation to compressed KV cache : Perform attention computation on the compressed representation, avoiding the overhead of expanding the cache and re-reading/writing VRAM.

TMA and pipeline scheduling : Use hardware asynchronous copy (TMA) to prefetch data and interleave data movement with matrix math, keeping GPUs computing rather than waiting.

Variable-length sequence and paged cache support : Handle different user context lengths and scattered cache blocks, adapting to multi-request concurrent inference.

FlashMLA's core logic is only a few hundred lines, making it the purest, most modern high-performance kernel textbook for developers learning modern GPU architecture, hardware pipelines, asynchronous memory copies, and assembly-level operator tuning.

6. DeepInfer

Created on October 16, reaching 18.6K stars in five days. This is the exact same underlying engine that powers hundreds of millions of daily production requests for DeepSeek. General-purpose inference frameworks like vLLM, SGLang, and TGI aim to support all model architectures and hardware, inevitably introducing heavy abstraction layers, scheduling glue code, and generic memory management.

DeepInfer takes a different approach: co-optimize operators, parallelism strategies, and scheduling around DeepSeek models and actual deployment hardware. Initially built on vLLM, key optimizations include:

Custom high-performance operators : Integrate FlashMLA, DeepGEMM, DeepEP to optimize attention, matrix ops, and expert communication respectively.

MoE expert parallelism and load balancing : Distribute experts across devices and adjust expert load to reduce stragglers.

Asynchronous scheduling : While GPU computes current batch, CPU prepares next batch, reducing scheduling wait.

Speculative decoding : Use DSpark to generate candidate tokens, then batch-verify with confidence-based verification scheduling to improve generation efficiency.

Comparing DeepInfer's execution pipeline with vLLM reveals how discarding the bulky general-purpose framework and tailoring every operator and scheduling stage to a specific model yields terrifying throughput and resource utilization.

5. DeepSeek-OCR

Pioneers contextual optical compression , compressing dozens of document pages into a few visual tokens. Traditional approaches OCR hundreds of pages of financial reports, dense papers, engineering drawings, or long codebases into plain text fed into the context window, suffering two fatal flaws:

Token explosion and compute surge : Hundreds of thousands of characters drastically inflate KV cache, causing quadratic attention cost explosion.

Catastrophic structural information loss : Table alignment, multi-column layouts, indentation, flowcharts, and image-text spatial relationships vanish when flattened into strings.

DeepSeek-OCR flips the paradigm: why convert to text at all? It renders documents as high-fidelity images, uses a dedicated Vision Encoder for patch-level fine-grained splitting and optical feature extraction, so a dense page containing thousands of text tokens is represented by only dozens to hundreds of tokens after optical compression.

4. DeepSeek-Coder

The foundational work for domestic coding LLMs, evolving into the world's first top-tier open-source MoE code flagship Coder-V2. The first generation launched in October 2023, marking many developers' first real encounter with DeepSeek.

In late 2023, open-source code models could barely handle single-line completion and failed at complex engineering; global developers relied solely on closed-source GPT-4. DeepSeek decided to pre-train a dedicated code base from scratch:

Pre-trained on 2 trillion tokens: 87% source code covering 80+ mainstream languages, 10% technical docs and bilingual code Q&A, 3% high-density math logic corpus.

Introduced bidirectional masked training with prefix, suffix, and middle infilling during pre-training, enabling not only forward continuation but also precise local refactoring and inline completion based on cursor context.

Supports 16K+ context, with targeted training on large-project dependency structures.

The June 2024 DeepSeek-Coder-V2 became the world's first open-source MoE code LLM (236B total / 21B active). It expanded language support to 338, context to 128K, and surpassed GPT-4-Turbo on HumanEval and SWE-bench, shattering the bias that coding requires foreign closed models.

3. DeepSeek-R1

The industrial revolution of LLM reinforcement learning and deep reasoning, breaking the OpenAI o1 monopoly myth. Released January 20, 2025; within a week its disruptive cost and compute restructuring effect triggered a capital market tsunami — Nvidia dropped 17% in a single day, wiping out nearly $590B. No open-source model had ever achieved this.

Before R1, reasoning models were widely believed to be an absolute moat only a few Silicon Valley giants could build using massive human-annotated chain-of-thought (CoT) and costly process reward models (PRM). High training costs and black-box algorithms made reasoning an ultra-expensive luxury.

The R1 technical report outlines a reinforcement-learning-centric reasoning training path with four key pillars:

R1-Zero explores reasoning via pure RL : Starting from DeepSeek-V3-Base without supervised fine-tuning. Math problems verified by answer checking, code problems by test execution; feedback trains the model to gradually exhibit self-checking and reflection behaviors.

GRPO reduces RL overhead : Model generates a group of answers per question; intra-group score comparison identifies better responses. This eliminates the separate value model typically required by PPO, cutting VRAM and compute during training.

R1 uses multi-stage training for usability : Small cold-start data regularizes output, then reasoning RL; then select high-quality answers for SFT; then further RL — balancing reasoning, readability, and general task performance.

Distillation lets small models learn reasoning : Fine-tune Qwen and Llama on R1-generated high-quality reasoning data, yielding distilled versions from 1.5B to 70B, lowering deployment barriers.

2. DeepSeek-V3

Released December 2024: 671B total parameters, 37B activated per token, trained on a 2048 H800 cluster for only ~$5.576M, matching and partially surpassing world-leading closed models on benchmarks.

Traditional Transformer architectures hit exponential cost walls, memory walls, and long-context KV cache explosions beyond hundreds of billions of parameters. DeepSeek-V3 pioneers feeding system-level hardware considerations back into algorithmic architecture design.

The V3 technical report highlights four must-remember innovations:

MLA: shrink KV cache — compact representation of history reduces inference VRAM, enabling long text and higher concurrency.

MoE with auxiliary-loss-free load balancing: invoke experts on demand — each token uses few experts; dynamic routing bias adjustment balances load, reducing waste from expert imbalance.

FP8 mixed-precision training: lower compute and storage overhead — most matmuls use compact FP8, sensitive parts keep higher precision, maintaining stability while boosting efficiency.

MTP multi-token prediction — training adds future token prediction objectives to improve learning; inference can use MTP module for speculative decoding to accelerate output.

The repository's inference/model.py is a masterclass: zero bloated third-party dependencies, just 800+ lines of extremely clean PyTorch implementing MLA, MoE dynamic routing, RMSNorm, RoPE crystal-clearly. Reading this single file thoroughly puts one's understanding of modern LLM engineering architecture ahead of the vast majority.

1. DeepSeek Harness

DeepSeek's most explosive 2026 open-source project — DeepSeek Harness. Built on the Cordis microkernel, a native next-gen agent runtime embracing everything-as-plugin and Harness Engineering.

Officially open-sourced under MIT on August 13, 2026; within two months stars surpassed 250K, becoming the most dominant agent infrastructure on GitHub and within the organization.

Many attribute DSH's explosion to DeepSeek's brand, but examining the .agents/notes/ directory in its repo reveals an engineering culture that commands awe. Under Cui Tianyi's team, every system evolution note carries real thinking dates, actual design alternatives faced, trade-offs under different hardware platforms, and abandoned technical routes. This panoramic, no-reservation disclosure of real system architecture design process is far scarcer and more震撼 than tens of thousands of lines of code.

Conclusion

Reviewing these 10 projects reveals DeepSeek's open-source trajectory differs fundamentally from other AI companies. Most vendors open-source a product and a trimmed inference demo for testing; DeepSeek open-sources a full-stack, top-down anatomical dissection:

At the operator layer, FlashMLA squeezes Hopper's TMA and memory bandwidth limits.

At the communication layer, DeepEP solves cross-node MoE expert parallel All-to-All throughput bottlenecks.

At the storage layer, 3FS leverages all-flash and RDMA to build the I/O aorta supporting 10K-GPU throughput.

At the model layer, DeepSeek-V3 and R1 prove MLA, auxiliary-loss-free MoE, and pure RL GRPO can beat expensive compute stacking.

At the serving layer, the production-grade DeepInfer inference engine is open-sourced without reservation.

At the application layer, DeepSeek Harness defines the next-gen agent harness engineering paradigm.

Study operators with FlashMLA, build clusters with DeepEP and 3FS, learn architecture with V3, master RL with Math and R1, run production with DeepInfer, engineer agents with Harness. No private tech, shared with the world. This composure in dissecting core productivity to the bone is what truly commands reverence for a top-tier technology company.