If you landed here asking whether DeepSeek V3 is still the model you should be building on in 2026, the short answer is: probably not for new projects, but understanding it matters more than ever. DeepSeek V3 was the December 2024 release that forced every major AI lab to re-examine training budgets, and its architecture still underpins the DeepSeek-R1 reasoning model and the V3.1 / V3.2 updates that followed. With DeepSeek V4 now shipping as the current generation, V3 has moved from frontier flagship to reference point — the model you study to understand how the MoE cost curve actually bends. This guide covers what V3 is, how it was built, what it scored, where it still has a role, and how to migrate off it cleanly.

What DeepSeek V3 is

DeepSeek V3 is an open-weight Mixture-of-Experts (MoE) language model from the Hangzhou-based lab DeepSeek, first released on December 27, 2024. It is a strong Mixture-of-Experts model with 671B total parameters with 37B activated for each token, adopting Multi-head Latent Attention (MLA) and DeepSeekMoE architectures. The release drew global attention not because of raw benchmarks alone, but because of the combination: frontier-competitive quality, open weights on Hugging Face, a detailed technical report, and a training bill that undercut Western labs by roughly two orders of magnitude.

V3 is no longer the current generation. DeepSeek V4 launched on April 24, 2026, shipping as two open-weight MoE tiers — deepseek-v4-pro (1.6T total / 49B active) and deepseek-v4-flash (284B / 13B active) — both under the MIT license with a 1M-token default context window. If you are choosing a model today, start at the DeepSeek V4 page. If you are auditing an existing integration that still references V3, keep reading — the rest of this article is for you.

Architecture and lineage

V3 sits at a specific point in the DeepSeek research line: Transformer → DeepSeek-V2 (May 2024) → DeepSeek-V3 (December 2024) → R1 (January 2025, built on the same base) → V3.1 (August 2025) → V3.2 (late 2025) → V4 (April 2026). Two architectural choices from V2 carry through V3 and are worth naming:

• Multi-head Latent Attention (MLA) — compresses the KV cache into a low-dimensional latent space, which is what makes long-context inference affordable at this scale.
• DeepSeekMoE — fine-grained expert specialisation with shared experts. Each MoE layer consists of 1 shared expert and 256 routed experts, where the intermediate hidden dimension of each expert is 2048, and among the routed experts, 8 experts are activated for each token.

V3 adds two things V2 did not have. First, an auxiliary-loss-free load-balancing strategy, which avoids the usual trade-off where the balancing loss degrades model quality. Second, a Multi-Token Prediction (MTP) training objective that predicts two tokens ahead during training and can later be repurposed for speculative decoding. DeepSeek pre-trained V3 on 14.8 trillion diverse and high-quality tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages, and the full training required only 2.788M H800 GPU hours.

Specifications at a glance

The 685B figure you may see on Hugging Face is not a contradiction — the total size of DeepSeek-V3 models on HuggingFace is 685B, which includes 671B of the main model weights and 14B of the Multi-Token Prediction (MTP) module weights.

Benchmarks from the V3 technical report

These are the headline numbers DeepSeek published alongside the December 2024 release. Every comparison here is against the specific competitor version DeepSeek listed, not against later revisions. If you are writing a 2026 comparison, use V4 benchmarks instead — see the DeepSeek benchmarks 2026 roundup.

Read those carefully. V3 led on math and competitive coding, matched Claude-3.5-Sonnet on general knowledge, and trailed Claude on GPQA Diamond. The R1 report later pushed MMLU to 90.8 for R1 while keeping V3 at 88.5, which is why R1 — not V3 — became the reasoning-workload default through 2025. For an independent re-test, see our DeepSeek V3 review.

Training cost: the $5.6M number in context

The figure that launched a thousand headlines: training required merely 2.788 million H800 GPU hours at a total cost of $5.576 million. Two things to hold steady here. First, that is training compute only — not research salaries, not data acquisition, not the cost of the prior V2 line that made V3 possible. Second, this $5.6M figure belongs to V3. The $294,000 figure you may have read belongs to R1 (disclosed by Reuters in September 2025). Do not blend them.

What the number changed was the public conversation about what frontier training has to cost. The DeepSeek funding and cost coverage holds more on the market reaction.

Strengths — where V3 still earned its keep

• Open weights at frontier scale. Before V3, “open-weight” and “GPT-4o-class” were not phrases that lived in the same sentence.
• Mathematics. 90.2 on MATH-500 in the base V3 report is a strong result for a non-reasoning model.
• Inference cost economics. Because the weights are public, multiple providers host V3 at aggressive per-token rates, which kept the DeepSeek API pricing below proprietary frontier rates throughout 2025.
• Foundation for R1. R1’s reasoning RL was applied on top of the V3 base. Without V3 there is no R1 — see DeepSeek R1.

Weaknesses — where V3 falls short today

• No native thinking mode. Original V3 is a non-reasoning chat model. V3.1 later merged thinking/non-thinking into one model; V3 did not.
• 128K context. Fine in 2024, middling in 2026. V4 defaults to 1M tokens.
• Hardware barrier for local inference. 671B parameters means serious metal. For self-hosters, see the DeepSeek hardware calculator.
• Weaker agentic / tool-use behaviour than the V3.1 and V4 generations, which were trained specifically for agent workflows.

How to access DeepSeek V3 today

Open weights

The original V3 and the March 2025 V3-0324 refresh are on Hugging Face under the deepseek-ai org. The V3-0324 model supports features such as function calling, JSON output, and FIM completion. If you want to run it offline, start with the install DeepSeek locally tutorial.

Hosted API — and the legacy-ID migration window

Here is where it gets sharp. On DeepSeek’s official API today, V3 is no longer a selectable model. The legacy IDs deepseek-chat and deepseek-reasoner still work, but they now route to deepseek-v4-flash (non-thinking and thinking respectively). Those legacy IDs are scheduled to be fully retired on 2026-07-24 at 15:59 UTC. After that date, requests using them will fail.

Chat requests hit POST /chat/completions, the OpenAI-compatible endpoint. The API is stateless — clients resend the full message history with every request — which is the opposite of how the web chat and mobile app behave, where session history is maintained for you. DeepSeek also ships an Anthropic-compatible surface against the same base URL, so the Anthropic SDK works with a base_url and api_key swap.

A minimal migration example in Python, using the OpenAI SDK:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.deepseek.com",
    api_key="YOUR_KEY",
)

resp = client.chat.completions.create(
    model="deepseek-v4-flash",  # was "deepseek-chat" on V3 / V3.x
    messages=[
        {"role": "system", "content": "You are a terse assistant."},
        {"role": "user", "content": "Summarise the V3 to V4 migration in one line."},
    ],
    temperature=1.3,  # DeepSeek's guidance for general conversation
    max_tokens=512,
)
print(resp.choices[0].message.content)

If you want thinking mode on V4, keep base_url, pass reasoning_effort="high", and set extra_body={"thinking": {"type": "enabled"}}. The response returns reasoning_content alongside the final content. Full walkthrough in the DeepSeek API getting started guide.

Pricing snapshot — and why V3’s old rates no longer apply

V3’s historical API rate card ($0.27 input / $1.10 output per million tokens) is retired. V4 replaced it. As of April 2026, the current DeepSeek rates are:

Worked example on V4-Flash — 1,000,000 calls with a 2,000-token cached system prompt, a 200-token user message, and a 300-token response:

• Cached input: 2,000 × 1,000,000 × $0.028 / 1M = $56.00
• Uncached input: 200 × 1,000,000 × $0.14 / 1M = $28.00
• Output: 300 × 1,000,000 × $0.28 / 1M = $84.00
• Total: $168.00

The off-peak discount that existed in 2025 ended on 2025-09-05 and has not returned. Rates on DeepSeek’s pricing page can change during a Preview window; verify at the DeepSeek API pricing reference before committing.

Best use cases — and honest alternatives

If you are running V3 today, the honest guidance is: migrate to V4-Flash for anything production, keep R1 on the shortlist for pure reasoning workloads, and only stay on V3 open weights if you have a specific reason (regulatory, air-gapped deployment, fine-tuning pipeline built around the V3 base).

• Code generation — V3 was strong at HumanEval; V4-Pro is stronger on SWE-Bench Verified. See DeepSeek for coding.
• Math-heavy workloads — V3 base and R1 both shine. For reasoning-first work, start with DeepSeek R1.
• Long-form writing and translation — V3 is competent at temperature 1.3; V4 is better at longer horizons because of the 1M-token context.
• Cost-sensitive chat apps — V4-Flash undercuts V3’s own 2025 rates.

V3 vs later DeepSeek models

For deeper head-to-heads, see DeepSeek V3 vs GPT-4o and DeepSeek V3.2. A broader catalogue is on the DeepSeek models hub.

Verdict

DeepSeek V3 was the release that made open-weight frontier models economically plausible, and it still deserves credit for that. In 2026 it is a reference architecture, not a deployment target. Migrate hosted workloads to V4-Flash, keep R1 in mind for reasoning, and hold on to V3 weights only if you have a concrete self-hosting or fine-tuning reason. The migration itself is a one-line model= swap — the base URL does not change.

Last verified: 2026-04-24. DeepSeek AI Guide is an independent resource and is not affiliated with DeepSeek or its parent company. Model IDs, pricing and API behaviour change; check the official DeepSeek documentation and pricing page before committing to a production decision.