DeepSeek V3 0324

DeepSeek V3 0324 is a refreshed release in the DeepSeek-V3 family, positioned as an iterative upgrade rather than a new architecture. According to the official Hugging Face model card, it builds on its predecessor and shows notable gains on standard reasoning benchmarks, including MMLU-Pro moving from 75.9 to 81.2, GPQA from 59.1 to 68.4, AIME from 39.6 to 59.4, and LiveCodeBench from 39.2 to 49.2. Those measured improvements frame V3 0324 as a meaningful step forward in mathematical reasoning and code generation while keeping the V3 design philosophy intact.

Beyond raw benchmark scores, the release emphasizes practical developer-facing refinements. The model card highlights improved code executability and more aesthetically polished web pages and game front-ends, alongside better Chinese writing style and content quality. As an open-weight release in the broader DeepSeek lineup, V3 0324 fits well for teams that want stronger reasoning and front-end coding behavior in a text-only assistant, sitting between the base V3 and later reasoning-focused variants in the family's evolution toward V3.1, V3.2, and V4.

Jiekou.AIdeepseek/deepseek-v3-0324deepseek

Open provider page Provider docs

Quick Info

Provider: Jiekou.AI
Model key: deepseek/deepseek-v3-0324
Release date: Jan 1, 2026
Last updated: Jan 1, 2026
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.28
Output token cost: $1.14

Limits

Output tokens: 163,840 tokens
Context window: 163,840 tokens

Transparent token rates

Compare DeepSeek V3 0324 pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about DeepSeek V3 0324

NovitaAI

CoverageAnalysis

RunPod's technical analysis (August 25, 2026) frames DeepSeek V3.1's changes against the V3-0324 baseline, explicitly naming that exact version multiple times. The analysis describes V3-0324 as demonstrating notable improvements over its predecessor DeepSeek-V3 while maintaining the same underlying model structure as t The article is a third-party retrospective where V3-0324 serves as the reference baseline for V3.1's architectural evolution rather than as the primary news subject. Published by RunPod — another serving provider — it discusses V3-0324's properties to contextualize V3.1's hybrid reasoning design, training-based behavio

Jiekou.AI

CoveragePreview

DeepSeek's official V4 Preview release (api-docs.deepseek.com/news/news260424, Apr 24, 2026) introduces two new open-weight MoE models with 1M default context: DeepSeek-V4-Pro (1.6T total / 49B active params) positioned as open-source SOTA on agentic coding with world-knowledge and reasoning rivaling top closed models, The API is live today — developers only need to swap the model name to deepseek-v4-pro or deepseek-v4-flash on the existing base URL, with support for both OpenAI ChatCompletions and Anthropic APIs, plus dual thinking/non-thinking modes for both models. Open weights are available on Hugging Face. The note explicitly wa