102: DeepSeek Launches Open Source Week—What Exactly Is Being Open-Sourced in Large Models?
Summary
DeepSeek’s Open Source Week shows that what it open-sources may extend beyond model weights to the AI infrastructure supporting online services. The first release, FlashMLA, is the first of five repositories over five days; DeepSeek describes these building blocks as “documented, deployed, and battle-tested.” The point behind “small but sincere” is to help the community better understand and reuse these efficiency tools.
王子涵 sees DeepSeek’s common technical thread as making every stage of training and inference more efficient. From MLA in V2, to predicting two tokens in parallel in V3, FP8 low-precision training, and shorter pipeline-parallel bubbles, to FlashMLA’s C++ operator optimization for KV Cache, his summary is: “I think it might really be every one of them,” preserving the uncertainty in his judgment.
What the industry calls an “open-source model” usually means only an open technical report and model weights; the genuinely scarce layers are inference frameworks, training frameworks, and datasets. The further down the stack, the higher the cost of整理 dependencies, standardizing code, and controlling sensitive information. Data is especially close to a red line: AllenAI and EleutherAI’s Pythia were cited as rare examples of deep openness, with the latter even releasing intermediate pretraining checkpoints and earning the description “nanny-style open source.”
Open source is an organizational capability that requires extra investment, not a matter of uploading internal code to GitHub. 王子涵 once spent 2–3 days migrating an internal implementation to a public framework and more than a week cleaning up the code. For researchers, “a week is a lot of time,” so the degree of openness directly reflects whether serving the community is a team priority.
For 王子涵, Issues and PRs are closer to real usage than Stars when measuring an open-source ecosystem. “I generally don’t look at how many Stars it has; I look at how many Issues it has”: adaptation requests, compilation errors, and bugs often show that users are actually running the code. PRs mean someone has proactively modified it, usually without direct material compensation. vLLM’s roughly 800 contributors and support for multiple hardware platforms show how an ecosystem can amplify one institution’s R&D investment.
Whether a company opens its strongest model or a smaller version ultimately depends on its monetization model and ecosystem goals. DeepSeek released its strongest model under the permissive MIT License. Companies that need to sustain subscription revenue or API premiums may retain their frontier versions while opening smaller models. 曼祺 noted that 文心 planned to open-source its model only on June 30; 王子涵认为 the workload may be an important reason, but the specifics still depend on its own plan.
The stronger the model, the sharper the conflict between full openness and controllable misuse. 王子涵’s Twitter account was targeted by what appeared to be a model-driven Agent using personalized invitations to steal accounts. He therefore suggested that in the future, “open a relatively small model while keeping the strongest one closed” may be more合理; otherwise everything from phishing to network abuse launched through 100 burner accounts could be scaled. On AI surpassing human intelligence, his answer was: “I brought a new player to the game. That is the meaning.”
Deep dive
1. 王子涵 Enters the Debate as Both User and Contributor
王子涵 is a first-year PhD student at Northwestern University’s MLL Lab, researching Large Language Model, Agent reasoning, and long-context efficiency. He interned at DeepSeek in 2024 and will spend the summer at Agent startup Utori.
He considers his first relatively influential open-source project to be an Agent benchmark launched in February 2023 and released that September. Many companies were already claiming to have built Agents, but there was no unified evaluation tool, so the team reorganized existing tasks into a format where models had to call external tools to solve them.
On DeepSeek-V2, he stressed that he was involved only in R&D, discussion, and experiments, “not a particularly core contributor.” He is now working on an open-source repo combining DeepSeek-R1 with Agent, studying what DeepSeek-R1’s pure-RL approach can achieve on different tasks when combined with multi-turn dynamics.
2. Open Source Week Starts with Online Infrastructure
DeepSeek previewed five repositories over five consecutive days, saying these “humble building blocks” had all been documented, deployed, and tested online. Its wording was “small but sincere,” accompanied by the statement that there was “no ivory tower, only pure garage spirit and community-driven innovation.”
The concurrently pinned Open Infra Index is organized around AI infrastructure, indicating that the week’s releases would involve infrastructure. The first release, FlashMLA, is an MLA inference optimization from the V2 framework.
王子涵 admitted that he had “no particular ability to predict this sort of thing.” He thought later releases might involve the training or inference frameworks for V3 or R1, but this was a statement of possibility, not a prediction about the remaining four releases.
3. DeepSeek’s Common Technical Thread Is Making Every Step More Efficient
王子涵 summarized V3’s focus in 3 directions: carrying forward Meta’s multi-token prediction, with the report describing 2 tokens being predicted in parallel; FP8 low-precision training, which stores numbers with fewer bits; and pipeline-parallel optimization.
Pipeline parallelism sends tensors through different parts of a model in sequence, but bubbles can emerge between forward and backward passes, leaving some layers idle. DeepSeek’s work further compresses these periods in which no vectors are being processed.
FlashMLA pushes efficiency down into Attention KV Cache and low-level operators. The code is written in C++, which in 王子涵’s view makes it essentially certain that the project contains systems-level optimizations rather than merely restating the same model computations.
From V2 to V3 to FlashMLA, he sees the same objective: efficiency. “I think it might really be every one of them”—all focused on efficiency and reducing training and inference costs. The low-precision work explicitly emphasizes saving compute while keeping training results broadly comparable.
4. Operator Optimization Means Crossing the Narrow Bridge Fewer Times—and Exposes Engineering Gaps
王子涵 used FlashAttention to explain operator fusion. If a large matrix must undergo 5 operations in sequence, it is better to move it into a smaller but faster compute layer once, perform all 5 operations, and move it back out once, rather than shuttling it from a larger but slower storage layer after every operation.
His metaphor was processing wheat by sending it across a narrow bridge. Repeatedly sending intermediate products back and forth means spending time crossing the bridge rather than processing the wheat. “Being able to complete all these operations in one pass” is the essence of the optimization.
When 曼祺 asked whether this was a difference in willingness or ability, 王子涵 clearly chose ability. He once tried to build online training but got stuck between 2 paths: the fastest inference frameworks freeze the model and cannot update it while generating, while rewriting a library that supports both high-speed inference and dynamic updates exceeded his team’s engineering capacity.
5. Indexing 5 Repositories Does Not Mean 5 Capabilities Have Already Merged
After FlashMLA was released, the community quickly asked for FP8 deployment code and an NPU version. 王子涵认为 that combining FP8 with MLA’s KV Cache optimization could make the system faster still. DeepSeek may have plans to combine its distributed technologies, but he left that as a “maybe.”
The Open Infra Index can solve discoverability, but not integration. Combining 5 libraries into a unified library whose capabilities can be freely called and layered would still require handling merge-related bugs. The community will actively reproduce the work, but who takes responsibility for integration depends on DeepSeek’s own plans.
6. README Is the Entry Point to a Repository; File Languages Reveal the Layer
王子涵 defines a repo as a “repository” for implementing an algorithm or concept. It can contain code, but people also use repos for assignments and even novels. Users can download the contents directly or pull them locally with git clone.
An organization’s homepage is useful for searching for the target repo, while language tags help with an initial read on implementation depth. FlashMLA is tagged C++, consistent with low-level operator optimization, but the language tag alone cannot fully explain a project’s purpose.
DeepSeek-R1’s repo has no programming-language tag because it currently carries mainly the paper explanation, model-running instructions, and a small number of commands rather than complete training code. 王子涵 believes an R1-related training framework may be opened later; for now, the community already has its own reproductions.
Once inside a repository, he starts with the README, using installation dependencies, benchmarks, or the first runnable command to build an overall picture. He then follows imports, function definitions, and call relationships layer by layer, learning the code base as if it were a tree.
7. Issues Are Closer to Real Usage Than Stars; PRs Mean Direct Co-Building
FlashMLA uses the MIT License, whose key features are permissiveness and “as is”: developers make the code available for free but do not guarantee the results of its use or accept responsibility for problems users encounter.
王子涵’s preferred metric is: “I generally don’t look at how many Stars it has; I look at how many Issues it has.” Stars are closer to a superficial like. Adaptation requests, compilation errors, and bugs in Issues often show that users have actually run the code. An Issue remains open until it is resolved or actively closed.
A pull request goes a level deeper: the contributor has modified the code and uses an added-and-deleted diff to show how the request was addressed, after which maintainers decide whether to merge it. External PRs usually offer no direct material reward. “The only motivation, I think, is passion”; GitHub badges provide only limited incentive.
8. An “Open-Source Model” Spans at Least 5 Layers; the Industry Usually Opens Only the Weights
王子涵 divides openness into 5 layers: technical report, model weights, inference framework, training framework, and dataset. Today, calling a model “open source” usually means only that a report exists and the weights can be downloaded; it does not automatically include everything needed to reproduce training.
The technical report is the most basic layer. It may contain no code and may describe the method ambiguously, but it should at least explain what approach was used. Model weights give users the result of training, but effective deployment still requires an inference framework.
Inference frameworks are opened less often, training frameworks even less often, and commercial companies almost never provide complete datasets. As a result, models carrying the same “open-source” label can differ dramatically in how understandable, modifiable, and reproducible they actually are.
9. V3’s Detailed Report Built Technical Trust Before R1 Took It Mainstream
王子涵 considers V3 one of the especially detailed reports among DeepSeek’s releases. The company’s technical team “all write code,” with individuals validating their own ideas and the successful ones entering the next training run. Dozens of small contributions ultimately came together in a report of more than 50 pages.
The contributor list includes not only 梁文锋 and technical staff, but also people involved in data annotation and business. As a base model, V3 covers the substantial work of pretraining, post-training, fine-tuning, and data; R1 focuses on explaining the reinforcement learning method built on the V3 base model.
曼祺’s view was that before R1 went viral with the general public, V3’s detailed experiments and ablation results had already earned respect globally, especially among the core technical community in the United States. 王子涵 agreed: even a ledger-like report should clearly record what was done and help readers understand the technology progressively, rather than leaving behind material that is difficult to use.
10. Hugging Face Delivers the Weights; vLLM and SGLang Expand Usability
Model weights are typically downloaded from Hugging Face. The model shown on the program was split into 163 shards, each potentially around 4GB. Users can load them through instructions provided by libraries such as Transformers or vLLM, with the framework handling the actual runtime.
vLLM originated in work by Berkeley researchers and later grew into an ecosystem with more than 800 contributors, supporting architectures including CPU, CUDA, NPU, and HPU. 王子涵 sees tens of thousands of Stars, thousands of Issues, and hundreds of PRs as evidence of deep usage.
After MiniMax-01 open-sourced a new linear-attention architecture, community feedback pointed to a lack of optimization, so its team contributed optimizations to vLLM. A mature framework can therefore absorb a model company’s deployment work without every company having to rebuild its own ecosystem.
SGLang is another fast-growing route. The usage experience 王子涵 has heard may be comparable to vLLM’s and better in some respects. Its team announced support for DeepSeek-V3 and R1 soon after R1’s release, showing a faster cadence in tracking new models.
11. Training Frameworks Show Most Clearly That Open Source Is Not “Dumping Internal Code”
ByteDance’s open-source verl supports reinforcement-learning training, including PPO trainer and FSDP parallelism modes, and has accumulated hundreds of Issues and PRs. 王子涵 used it to complete research and said bluntly that this openness gave him “a pretty good impression” of ByteDance.
He also introduced ESFT, his own work, which uses different experts’ task affinity in MoE models to fine-tune only the relevant experts while freezing the rest. This reduces compute requirements and embodies the principle of “letting specialists do specialist work.”
The difficulty was that the original implementation depended on a large internal company framework that was not yet planned for release. He had to rewrite the relevant components on a public framework, spending 2–3 days on adaptation and more than a week standardizing the code. “A week is a lot of time.”
Standards also do not emerge overnight. He has seen projects name 3 files step 1, step 2, and step 3; if a later contributor inserts a step, it can only be called step 1.5 or require renaming every file. Clear naming and consistent formatting determine whether others can genuinely maintain and extend the code.
12. Data Remains the Deepest Red Line; Privacy Risks Outweigh Reproduction Value
Commercial companies almost never release complete training datasets, at most disclosing the proportions of different data sources in a report. 王子涵认为 the reason is not simply that they “don’t want to open it,” but also the sensitivity of the information and the security responsibilities involved.
Even when a webpage is public and scraping it is legally permitted, consolidating large numbers of personal websites into a dataset can allow users to locate personal information quickly and use it for harmful purposes. Information that was dispersed and visible becomes centralized and searchable, changing the risk profile.
The few exceptions he cited include AllenAI and EleutherAI. The latter’s Pythia not only releases the related work, but also provides pretraining checkpoints at steps 0, 1, 2, and 4 through step 1,000, then every 1,000 steps thereafter. 曼祺因此 described it as “nanny-style open source.”
13. Opening the Strongest Model or a Smaller One Ultimately Depends on Monetization and Ecosystem Goals
DeepSeek has long released its strongest models under the permissive MIT License. 曼祺 also noted that MiniMax and 阶跃星辰 began releasing their first open-source models in January, and that 文心 announced a formal open-source release by June 30.
When 曼祺 asked why 文心 would wait so long, 王子涵认为 the workload was certainly an important reason, but still said it “depends on their own plan.” Companies without an accumulated open-source foundation need to invest additional work in code standards, dependency replacement, documentation, and maintenance staff.
王子涵 links the strategy to monetization. Institutions that do not depend on model subscriptions or API premiums may open their strongest models and even seek to reshape the industry ecosystem, making their system a standard. Companies that need to preserve a commercial lead may open smaller versions to serve the community while retaining their core advantage.
Asked whether OpenAI would open its strongest model, he answered that he was “not particularly sure.” Sam Altman’s options in the vote at the time were o3-mini or a phone-size model deployable on mobile devices. 曼祺 then argued that either choice could serve some users, while 王子涵 turned to the possibility that keeping a model closed could also be rational.
14. The Stronger the Model, the Harder It Is to Avoid the Tension Between Openness and Misuse
王子涵’s Twitter account was hacked. He suspects the attacker used a Large Language Model to create an Agent that generated different versions of a business-interview invitation for different users, ultimately诱cing them to click a link. The attacker may even have continued training the model to improve its click-through rate.
Based on this, he suggested that once models become powerful enough, it may be more rational to “open a relatively small model while keeping the strongest one closed,” because developers will find it difficult to stop others from using it for malicious purposes.
His escalation scenario is not abstract doomsday theory but cheaper, scalable harm. The strongest model could do more than phishing; it could also automatically register 100 burner accounts and launch coordinated online abuse. 王子涵 confirmed that this view is not an argument for commercial protection, but a consideration of technology’s social impact.
Faced with the prospect of AI surpassing human intelligence, 王子涵 also rejected the idea that intelligence is the only form of value. The Go world moved from resisting AI to learning from it and enjoying it: “Then there’s no need to compete anymore”; “just enjoy it.” Humans enter the world like registering a game account: “I brought a new player to the game. That is the meaning.”