Official Technical Resources: DeepSeek
Official Technical Resources · DeepSeek. Technical reports, model cards and API changelogs. This source has 10 entries admitted under this category's review policy; each links directly to the original.
Authority tiers in this category: A1–A2.
DeepSeek-V4-Pro model card
The official open weights and model card for the Pro tier of the V4 series, a text generation model also offered as an API service. The model card gives the architecture, usage and license terms, and is the primary source for checking whether the open-weight release and the online API version are the same tier.
Evidence use: Directly supports statements such as "whether V4-Pro is open source, under what license, and which tasks the vendor says it supports"; cross-model capability comparisons need independent evaluations.
Limitations: Model cards are self-reported by the publisher; capability claims need cross-checking through third-party evaluation or independent reproduction. Weight licenses and commercial use terms are governed by the model card and accompanying license, and may change at any time.
Related concepts: Large Language Model (LLM) · Major Model Families · Pre-training
Link last checked: 2026-09-23
DeepSeek-V4.1-Flash release announcement
The official announcement of V4.1-Flash, framed as "smarter, faster, cheaper", which positions this version on the speed and cost side. It is the model's first public statement of capability and efficiency and should be read alongside the pricing page and actual API behavior.
Evidence use: Supports statements such as "when this version launched and how the vendor positions it"; specific performance and price figures must be checked against the pricing page or actual tests.
Limitations: Announcements are self-reported by the publisher; performance and price figures should be judged by actual API behavior and billing. Announcements are not revised after publication; cite them with their publication date and do not treat them as current status.
Related concepts: Large Language Model (LLM) · Model Selection and Cost · LLM Inference Optimization
Link last checked: 2026-09-23
DeepSeek V4 Preview release announcement
The launch announcement for the V4 preview, which states that it "enters the era of cost-effective 1-million-token context" and announces open-sourcing. It marks the start of the V4 series (including the later Pro and Flash) and is the original source for the claim that million-token context has entered a low price tier.
Evidence use: Supports statements such as "when long context entered a lower price tier and who first announced it as open source"; actual context limits and billing still need checking on the current pricing page.
Limitations: Announcements are self-reported by the publisher; performance and price figures should be judged by actual API behavior and billing. Announcements are not revised after publication; cite them with their publication date and do not treat them as current status.
Related concepts: Context Window · Large Language Model (LLM) · Major Model Families
Link last checked: 2026-09-23
DeepSeek Harness (agent runtime)
The officially open-sourced agent runtime framework, self-described as "Everything is a Plugin". Written in TypeScript and tagged AI Agent and plugin system, with its home page at deepseek.com/harness. It is the most-watched of the organization's public repositories.
Evidence use: Supports statements such as "whether the vendor offers its own agent runtime and how it is extended"; the merits of the framework need separate testing.
Limitations: The repository is maintained by the vendor; the permitted use of code and weights is governed by the license declared in the repository. Star counts and activity are only signals of attention, not evidence of capability.
Related concepts: AI Agent · Agent Frameworks · Agent Skills
Link last checked: 2026-09-23
DeepSeek-V3 official repository
The official repository of the third-generation flagship model, with weights and guidance for inference and fine-tuning. The series' published architecture (multi-head latent attention combined with mixture of experts) was carried into the V4 series and was the starting point for community-written inference kernels.
Evidence use: Supports statements such as "when the architecture was open-sourced and which deployment paths are officially provided"; the architecture's effectiveness should be judged from the paper or independent reproduction.
Limitations: The repository is maintained by the vendor; the permitted use of code and weights is governed by the license declared in the repository. Star counts and activity are only signals of attention, not evidence of capability.
Related concepts: Mixture of Experts (MoE) · Attention · Large Language Model (LLM)
Link last checked: 2026-09-23
DeepSeek-R1 official repository
The official repository of the R1 reasoning model, describing its training approach of eliciting long-chain reasoning with reinforcement learning and providing distilled smaller models. The series is a representative output that brought chain of thought into public discussion and released weights at scale.
Evidence use: Supports as the original source for discussions such as "whether reasoning ability can be elicited from a base model with reinforcement learning"; the strength of the conclusion depends on later reproductions.
Limitations: The repository is maintained by the vendor; the permitted use of code and weights is governed by the license declared in the repository. Star counts and activity are only signals of attention, not evidence of capability.
Related concepts: Reasoning Models · Chain of Thought (CoT) · RLHF and Preference Alignment
Link last checked: 2026-09-23
DeepSeek-OCR official repository
The official OCR model repository, focused on "contexts optical compression": compressing document images into fewer vision tokens before handing them to a language model. It represents a route for long documents different from plain text chunking and is primary material for understanding the relationship between multimodality and context cost.
Evidence use: Supports statements such as "whether visual compression can reduce token usage for long documents"; the trade-off between compression ratio and accuracy should be judged from the repository's experimental data.
Limitations: The repository is maintained by the vendor; the permitted use of code and weights is governed by the license declared in the repository. Star counts and activity are only signals of attention, not evidence of capability.
Related concepts: Multimodal Models · Context Compaction · Tokens and Tokenization
Link last checked: 2026-09-23
Thinking Mode
Explains the reasoning mode in which the model outputs a chain of thought before the final answer, and how request parameters and response fields differ in this mode. It is the official line on how reasoning models are distinguished from ordinary chat models in engineering terms.
Evidence use: Directly supports engineering statements such as "how to enable reasoning for this model and how the response structure differs".
Limitations: Supports only the vendor's current public interfaces, parameters and official recommendations; it cannot replace independent cross-platform comparison. Online documentation changes over time; recheck the original page for versions, pricing, limits and regional availability.
Related concepts: Reasoning Models · Chain of Thought (CoT) · Sampling and Decoding Parameters
Link last checked: 2026-09-23
Context Caching
Introduces on-disk context caching, enabled by default for all users: cached input tokens are billed at a lower price, with no change to how the API is called. It turns prompt caching from an optional optimization into default behavior, and is key to understanding how the pricing structure relates to the caching mechanism.
Evidence use: Supports statements such as "how caching affects cost and whether code changes are needed"; hit rates and actual discounts should follow the bill.
Limitations: Supports only the vendor's current public interfaces, parameters and official recommendations; it cannot replace independent cross-platform comparison. Online documentation changes over time; recheck the original page for versions, pricing, limits and regional availability.
Related concepts: Prompt Caching · LLM Inference Optimization
Link last checked: 2026-09-23
Models & Pricing
Lists input, output and cache-hit prices per million tokens for each model; the original price table for judging its cost-effectiveness. Prices change over time, so citations must give the date read and be checked against actual bills.
Evidence use: Directly supports statements such as "the official list price of a model", but the date it was read must be given.
Limitations: Supports only the vendor's current public interfaces, parameters and official recommendations; it cannot replace independent cross-platform comparison. Online documentation changes over time; recheck the original page for versions, pricing, limits and regional availability.
Related concepts: Model Selection and Cost · Prompt Caching
Link last checked: 2026-09-23