DeepSeek V2 Explained: The MoE Model That Started a Lineage

Why write about DeepSeek V2 in 2026, several model generations after it shipped? Because V2 is where Multi-head Latent Attention (MLA) first appeared and where DeepSeek’s MoE design was first scaled to a flagship model; V3, R1 and V3.2 kept both ideas, and their published configs use the same latent-KV rank (512) as V2. If you want to understand where MLA came from, or why DeepSeek’s later models are built the way they are, V2 is the place to start.

This article gives you the full V2 picture: parameter counts, the benchmarks DeepSeek published in the original technical report, what V2 was good and bad at, how to access the open weights today, and how V2 relates to the models DeepSeek’s API serves now. By the end you’ll know whether V2 is still worth your time.

What is DeepSeek V2?

DeepSeek V2 is an open-weight Mixture-of-Experts (MoE) language model from DeepSeek-AI. The GitHub README dates its release to 6 May 2024, and the technical report was posted to arXiv on 7 May 2024. In the model card’s words, it is a strong MoE language model characterised by economical training and efficient inference, with 236B total parameters, of which 21B are activated for each token, and it supports a 128K-token context. The release moved DeepSeek’s flagship line from dense models (the earlier 67B) to sparse MoE, building on the smaller DeepSeekMoE 16B research model from January 2024.

Two ideas matter most. The first is Multi-head Latent Attention (MLA), which compresses the keys and values that drive attention into a much smaller cache. The second is DeepSeekMoE, a fine-grained expert-routing scheme with shared experts. Together they made V2 cheap to train and fast to serve: compared with DeepSeek 67B, DeepSeek reports that V2 saves 42.5% of training costs, reduces the KV cache by 93.3% and boosts maximum generation throughput to 5.76 times.

Architecture and lineage

Parameter count and context

• Total parameters: 236B
• Active parameters per token: 21B
• Context length: 128K tokens
• Pre-training corpus: 8.1 trillion tokens
• Release date: 6 May 2024 (GitHub README); arXiv report 7 May 2024
• Variants: DeepSeek-V2 (base) and DeepSeek-V2-Chat (RL) at launch, with DeepSeek-V2 Chat (SFT) also reported in the paper; DeepSeek-V2-Lite (16B total / 2.4B active, 32K context) followed on 16 May 2024, and a DeepSeek-V2-Chat-0628 checkpoint was published later

According to the paper, compared with the corpus used for DeepSeek 67B, the 8.1T-token corpus features an extended amount of data, especially Chinese data, and higher data quality. The Chinese data weighting is one reason V2 scores well on Chinese-language evaluations relative to its size.

Multi-head Latent Attention (MLA)

Standard multi-head attention stores a separate key and value vector for every token, every layer and every head, so at long context lengths the KV cache dominates GPU memory. The core of MLA, per the V2 paper, is low-rank joint compression of keys and values: the model caches a small latent vector per token (a KV compression dimension of 512 in the published config) instead of full keys and values. During inference the up-projection matrices can be absorbed into the query and output projections, so keys and values never need to be computed out in full; a separate, decoupled path carries the rotary position embedding.

This is what made a 128K context economical for V2 in 2024. DeepSeek-V3’s model card says V3 adopts MLA and DeepSeekMoE, “which were thoroughly validated in DeepSeek-V2”.

DeepSeekMoE

For FFNs, DeepSeek-V2 employs the DeepSeekMoE architecture, which has two key ideas: segmenting experts into finer granularity for higher expert specialisation and more accurate knowledge acquisition, and isolating some shared experts to mitigate knowledge redundancy. In V2’s published config, each MoE layer has 2 shared experts and 160 routed experts, 6 of which are activated per token. The paper also describes device-limited routing, which sends each token to experts on at most 3 devices to bound communication cost, part of why DeepSeek’s training-cost numbers came out so low.

Lineage to current models

DeepSeek’s next generations built directly on V2. DeepSeek-V3 is an MoE model with 671B total parameters and 37B activated per token that adopts MLA and DeepSeekMoE. R1 was trained on the V3 base; V3.1 added a hybrid thinking mode; V3.2 added DeepSeek Sparse Attention. The V4 generation is still MoE but brought new attention designs: hybrid Compressed Sparse Attention and Heavily Compressed Attention in the April 2026 preview, and CSA2 in V4.1-Flash. DeepSeek’s API today serves deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro (DeepSeek-V4-Pro-0813).

For the live picture of where this lineage now sits, see DeepSeek V4 and the DeepSeek models hub.

DeepSeek V2 benchmarks

The numbers below are DeepSeek-reported, from the original V2 technical report (arXiv 2405.04434, May 2024) and the model card. Comparison rows reference the model versions DeepSeek cited at the time, primarily LLaMA 3 70B Instruct and Mixtral 8x22B. Do not read these as competitive against current-generation models; for newer comparisons see the DeepSeek benchmarks 2026 roundup.

Per the V2 paper, DeepSeek-V2 Chat (RL) achieves a 38.9 length-controlled win rate on AlpacaEval 2.0, 8.97 overall on MT-Bench and 7.91 on AlignBench. On MMLU, the base model scored 78.5 against LLaMA3 70B’s 78.9 with only 21B activated parameters; the headline framing was efficiency per parameter, not absolute capability.

One thing the report is careful about: DeepSeek-V2 Chat (SFT) shows similar performance to LLaMA3 70B Chat in code and maths benchmarks, LLaMA3 70B Chat is better on MMLU and IFEval, and DeepSeek-V2 Chat (SFT) is stronger on Chinese tasks. V2 was not a sweep; it was a careful trade-off. DeepSeek-V2 Chat (RL) further improves on the SFT model in maths and coding.

Strengths — where V2 specifically won

• Efficiency per active parameter. Activating only 21B of 236B total parameters made V2 much cheaper to run per token than a dense model of similar quality; DeepSeek reports 5.76 times the maximum generation throughput of its dense 67B.
• KV cache footprint. The 93.3% KV-cache reduction versus the dense 67B baseline means far more concurrent long-context conversations fit in the same GPU memory.
• Maths and code (Chat RL). The RL-tuned chat variant reached 92.2 on GSM8K and 81.1 on HumanEval, strong numbers for an open-weight model in mid-2024.
• Chinese-language tasks. The base model scored 81.7 on C-Eval and 84.0 on CMMLU against LLaMA3 70B’s 67.5 and 69.3 in DeepSeek’s table.
• Open weights with a commercial-use licence. The code repository is MIT-licensed; use of the model weights is subject to DeepSeek’s separate Model License, and DeepSeek states that the V2 series supports commercial use. Read the licence, which includes use restrictions, before redistributing.

Weaknesses — where V2 fell short

• Absolute MMLU. V2 Chat (RL)’s 77.8 trailed LLaMA3 70B Instruct’s 80.3 in DeepSeek’s own table, and the paper concedes that V2 Chat’s reasoning still lagged larger models such as GPT-4 on AlignBench.
• Instruction following and IFEval. The V2 paper itself flags that LLaMA3 70B Chat was stronger on IFEval. DeepSeek’s Change Log later reported instruction-following gains for the API’s V2-0517 update (May 2024) and for DeepSeek V2.5 (September 2024).
• Hardware to self-host. DeepSeek’s README says BF16 inference with the full model needs 8 × 80GB GPUs; the current Hugging Face files for DeepSeek-V2 total about 471.5 GB. DeepSeek-V2-Lite (about 31.4 GB) is the smaller option, and its card says one 40 GB GPU is enough for BF16 inference. For sizing, the DeepSeek hardware calculator is useful.
• No reasoning mode. V2 predates DeepSeek’s thinking-mode models: there is no reasoning_content output. For that you want R1 or one of today’s API models, where thinking mode is on by default.
• Superseded. DeepSeek has replaced V2 with several newer generations; it is a 2024 model and its benchmarks reflect the field as it stood then.

How to access DeepSeek V2 today

V2 is no longer the default model anywhere. These are the access paths:

1. Open weights on Hugging Face. The deepseek-ai/DeepSeek-V2, DeepSeek-V2-Chat, DeepSeek-V2-Lite and DeepSeek-V2-Lite-Chat repositories remain online for self-hosting and research; DeepSeek’s README recommends vLLM or SGLang over plain Transformers for speed. See the install DeepSeek locally tutorial for the practical setup.
2. Not on the API. DeepSeek’s Change Log shows the deepseek-chat API model running V2 checkpoints in 2024 (upgraded to DeepSeek-V2-0517 on 17 May and to DeepSeek-V2-0628 on 28 June), then V2.5 from 5 September 2024 and V3 from 26 December 2024. The deepseek-chat and deepseek-reasoner IDs were scheduled for retirement on 2026-07-24 at 15:59 UTC and are no longer listed. Today’s pricing page (checked 26 October 2026) lists only deepseek-flash and deepseek-v4-pro, neither of which is V2.
3. Self-hosted only for V2-specific work. If you specifically need V2 weights, for reproducibility of a 2024 paper for example, Hugging Face is the only option.

Calling the current API instead

If you came here because an old tutorial pointed you at V2 and you just want a working chat call, use one of the current models: deepseek-flash or deepseek-v4-pro. Chat requests go to POST /chat/completions, the OpenAI-compatible endpoint at https://api.deepseek.com; an Anthropic-compatible endpoint is at https://api.deepseek.com/anthropic. The API is stateless, so your client must resend the conversation history with each request. Thinking mode is on by default and returns reasoning_content alongside the final content; V2 has no equivalent. For code samples, see the DeepSeek API getting started guide.

Pricing snapshot — V2’s API history and what current models cost

V2’s 2024 API prices are not on DeepSeek’s current pricing page, and V2 has no API price today. Behind deepseek-chat, the Change Log records V2-0517 and V2-0628 in 2024, then V2.5, V2.5-1210 and V3. As of October 2026, the rates for the current models per 1M tokens are:

Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays; all other hours are off-peak at half price. Always verify on the DeepSeek API pricing page before committing to a budget.

Best use cases for V2 (in 2026)

Very few. V2 is now a research artefact more than a working tool. The cases where it still makes sense:

• Reproducing the 2024 paper for academic work: use the original Hugging Face checkpoint.
• Studying MLA implementations: V2 is the model that introduced MLA, and its modelling code is published with the Hugging Face weights.
• Air-gapped Chinese-language deployment on hardware that can host 236B but not 671B: in DeepSeek’s 2024 tables, V2 scored well above LLaMA3 70B and Mixtral 8x22B on C-Eval and CMMLU.

For everything else (coding, writing, agents), use DeepSeek’s current models. See DeepSeek for coding and DeepSeek for developers for current-generation workflows.

Comparable alternatives

If you are comparing V2 historically, the obvious peers are LLaMA 3 70B and Mixtral 8x22B. If you are comparing the V2 line to a current alternative, the more useful pages are:

• DeepSeek V2.5: the September 2024 model that merged V2 Chat and Coder V2, with instruction-following and writing improvements per the Change Log
• DeepSeek V3: the 671B/37B successor that took over deepseek-chat on 26 December 2024
• DeepSeek vs Llama: the open-weight comparison most V2 readers want
• Open-source AI like DeepSeek: broader landscape

Verdict

DeepSeek V2 is the model where DeepSeek introduced MLA and scaled DeepSeekMoE to a flagship, and V3, R1 and V3.2 kept both designs. As a working tool in 2026 it is obsolete; as an architectural reference it is essential reading. If you are choosing a model to use today, pick one of DeepSeek’s current API models (deepseek-flash is the default recommendation). If you want to understand why DeepSeek’s models are built the way they are, read the V2 technical report first.