DeepSeek V3 Explained: The 671B MoE That Reset Open-Weight Economics

If you landed here asking whether DeepSeek V3 is still the model you should be building on in 2026, the short answer is: not for new projects, but understanding it still matters. DeepSeek V3 was the December 2024 release whose reported training cost drew worldwide attention, and its architecture underpins the DeepSeek-R1 reasoning model and the V3.1 and V3.2 updates that followed. DeepSeek’s API now serves deepseek-flash and deepseek-v4-pro, so V3 has moved from flagship to reference point — the model you study to understand how the MoE cost curve 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, announced on 26 December 2024 (DeepSeek’s release note and API Change Log both carry that date). Per its model card, it is “a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token”, adopting Multi-head Latent Attention (MLA) and DeepSeekMoE architectures. The release drew attention for the combination: benchmark results competitive with leading closed models of the time, open weights on Hugging Face, a detailed technical report and a reported training compute bill of $5.576M.

V3 is no longer the current generation. DeepSeek’s API today lists two models: deepseek-flash, served by DeepSeek-V4.1-Flash-1017, and deepseek-v4-pro, served by DeepSeek-V4-Pro-0813, both with MIT-licensed open weights, a 1M-token context and up to 384K output tokens. 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.

Architecture and lineage

V3 sits at a specific point in the DeepSeek research line: DeepSeek-V2 (2024) → DeepSeek-V3 (December 2024) → R1 (January 2025, built on the V3 base) → V3-0324 (March 2025) → V3.1 (August 2025) → V3.2 (December 2025) → the V4 preview (April 2026) → today’s V4-Pro-0813 and V4.1-Flash-1017. 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. Per the V3 report, each MoE layer consists of 1 shared expert and 256 routed experts with an intermediate hidden dimension of 2048 each, and 8 routed 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: with the prediction depth set to 1, each token also predicts one additional token beyond the next, and the module can be reused for speculative decoding. DeepSeek pre-trained V3 on 14.8 trillion tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages, and reports that 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 model card notes that the total size of DeepSeek-V3 models on Hugging Face is 685B, which includes 671B of main model weights and 14B of Multi-Token Prediction (MTP) module weights.

Benchmarks from the V3 technical report

These are the headline chat-model numbers DeepSeek published with the December 2024 release (V3 model card and technical report; DeepSeek-reported). Every comparison here is against the specific competitor version DeepSeek listed, not against later revisions. For current-model results, see the DeepSeek benchmarks 2026 roundup.

Read those carefully. On DeepSeek’s numbers, V3 led on math and competitive coding, roughly matched Claude-3.5-Sonnet on MMLU, and trailed Claude on MMLU-Pro and GPQA Diamond. DeepSeek’s later R1 report lists R1 at 90.8 on MMLU against V3’s 88.5, with much larger gains on reasoning-heavy math such as AIME 2024 (79.8 vs 39.2). See also our DeepSeek V3 review.

Training cost: the $5.6M number in context

The figure that launched a thousand headlines comes from DeepSeek’s V3 technical report: 2.788 million H800 GPU hours — 2,664K for pre-training, 119K for context-length extension and 5K for post-training — which, at an assumed rental price of $2 per GPU hour, comes to $5.576 million; pre-training ran on a cluster of 2,048 H800 GPUs in under two months. Two things to hold steady here. First, DeepSeek itself notes that this covers only the official training of V3, “excluding the costs associated with prior research and ablation experiments” — and it is not salaries, data acquisition or hardware ownership either. Second, this $5.6M figure belongs to V3. The $294,000 figure you may have read belongs to R1’s reasoning training, from the supplementary information of DeepSeek’s September 2025 Nature paper. 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. V3’s own results table puts it ahead of GPT-4o-0513 on most of the benchmarks listed, from a model anyone could download.
• Mathematics. 90.2 on MATH-500 in the V3 report is a strong result for a non-reasoning model.
• Low API prices at the time. From 8 February 2025, DeepSeek charged $0.27 per million cache-miss input tokens, $0.07 on a cache hit and $1.10 per million output tokens for V3 (see the pricing snapshot below).
• 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 and non-thinking into one model; V3 did not.
• 128K context. Fine in 2024, modest in 2026: today’s deepseek-flash and deepseek-v4-pro offer 1M tokens.
• Hardware barrier for local inference. 671B parameters means serious hardware. For self-hosters, see the DeepSeek hardware calculator.
• Weaker agentic / tool-use behaviour than later generations: DeepSeek’s Change Log credits V3.1 (August 2025) with “major improvements in tool usage and intelligent agent tasks”.

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; each repository’s current files total about 688.6 GB (checked 26 October 2026), and DeepSeek provides FP8 weights with a script to convert them to BF16. Per its model card, V3-0324 supports function calling, JSON output and FIM completion. On Ollama, the deepseek-v3:671b tag is listed at 404 GB (checked the same day). If you want to run it offline, start with the install DeepSeek locally tutorial.

Hosted API — what happened to deepseek-chat

On DeepSeek’s API, V3 is no longer served. The Change Log traces the deepseek-chat name: it was upgraded to DeepSeek-V3 on 26 December 2024 and to DeepSeek-V3-0324 on 24 March 2025. On 21 August 2025 it became the non-thinking mode of DeepSeek-V3.1 (with deepseek-reasoner, which had carried R1, becoming V3.1’s thinking mode), and it then followed V3.1-Terminus, V3.2-Exp and V3.2. On 24 April 2026 both names were pointed to deepseek-v4-flash (a legacy name now served by DeepSeek-V4.1-Flash-1017), and they were scheduled for retirement on 2026-07-24 at 15:59 UTC; they are no longer listed in DeepSeek’s docs.

Chat requests hit POST /chat/completions, the OpenAI-compatible endpoint at https://api.deepseek.com. The API is stateless — clients resend the full message history with every request. DeepSeek also accepts the Anthropic API format at a separate base URL, https://api.deepseek.com/anthropic. To migrate V3-era code, change the model field to deepseek-flash (or deepseek-v4-pro); the base URL does not change. Thinking mode is on by default on both models, so pass extra_body={"thinking": {"type": "disabled"}} if you want the old non-thinking deepseek-chat behaviour. Full walkthrough in the DeepSeek API getting started guide.

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

V3’s API rates are history. DeepSeek’s 26 December 2024 announcement kept V2 pricing until 8 February 2025, then set V3 at $0.27 per million input tokens on a cache miss, $0.07 on a cache hit and $1.10 per million output tokens. Wayback Machine captures of DeepSeek’s pricing page show that from 26 February 2025 until 16:00 UTC on 5 September 2025 these were the standard rates, with a 50% discount for V3 between 16:30 and 00:30 UTC ($0.135 cache miss, $0.035 cache hit, $0.55 output). Today’s rates (DeepSeek pricing page, checked 26 October 2026), in USD per 1M tokens:

Off-peak rates are half of peak; peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. Worked example — 1,000,000 calls with a 2,000-token cached system prompt, a 200-token user message and a 300-token response (2,000M cache-hit, 200M cache-miss and 300M output tokens):

• V3 at its 2025 standard rates: 2,000 × $0.07 = $140.00; 200 × $0.27 = $54.00; 300 × $1.10 = $330.00; total $524.00
• V3 at its 2025 discount rates (16:30–00:30 UTC, 26 Feb – 5 Sep 2025): 2,000 × $0.035 = $70.00; 200 × $0.135 = $27.00; 300 × $0.55 = $165.00; total $262.00
• deepseek-flash, peak: 2,000 × $0.006 = $12.00; 200 × $0.30 = $60.00; 300 × $1.20 = $360.00; total $432.00
• deepseek-flash, off-peak: 2,000 × $0.003 = $6.00; 200 × $0.15 = $30.00; 300 × $0.60 = $180.00; total $216.00

Note that deepseek-flash costs slightly more per token than V3 did on cache-miss input and output (peak $0.30 / $1.20 against V3’s standard $0.27 / $1.10; off-peak $0.15 / $0.60 against V3’s discount $0.135 / $0.55); the saving in this example comes from its much lower cache-hit rate. Verify current figures in our DeepSeek API pricing reference before committing.

Best use cases — and honest alternatives

If you are running V3 today, the honest guidance is: move hosted workloads to deepseek-flash, consider R1’s open weights for self-hosted reasoning, 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 scored 82.6 on HumanEval-Mul and 42.0 on SWE-Bench Verified in its own report; for hosted coding today use the current models. See DeepSeek for coding.
• Math-heavy workloads — V3 and R1 both shine. For reasoning-first work, start with DeepSeek R1.
• Long-form writing and translation — V3 is competent; today’s models add a 1M-token context for longer documents.
• Cost-sensitive chat apps — compare V3’s self-hosting costs with deepseek-flash’s peak and off-peak rates above.

V3 vs later DeepSeek models

* Preview-only history: the parameter counts DeepSeek published for the April 2026 V4 preview models (V4-Pro-Preview 1.6T total / 49B active, V4-Flash-Preview 284B / 13B) describe those preview releases only and are not stated here for the models the API serves today.

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 look economically plausible, and it still deserves credit for that. In 2026 it is a reference architecture, not a deployment target. Move hosted workloads to deepseek-flash (or deepseek-v4-pro), keep R1 in mind for self-hosted 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= change — the base URL does not change.