Skip to content

Sync latest model docs from BAI-docs #47

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Will-Guan merged 3 commits into from
Apr 28, 2026
Merged

Sync latest model docs from BAI-docs #47

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
c198cb6
Sync new model docs from BAI-docs
Apr 27, 2026
4a1c769
update
jizhen181-dot Apr 27, 2026
85e3d1b
Merge branch 'main' into ai-bankofai-patch-1
jizhen181-dot Apr 28, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
43 changes: 43 additions & 0 deletions docs/llmservice/models/deepseek-v4-flash.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,43 @@
# DeepSeek V4 Flash

## Overview

DeepSeek-V4-Flash is DeepSeek's high-efficiency open-source language model, released alongside V4-Pro on April 24, 2026 under the MIT License. With 284 billion total parameters and only 13 billion active parameters, it delivers performance within striking distance of V4-Pro at 12.4x lower cost, making it one of the most cost-effective models available.

## Key Features

* **Ultra-Efficient Architecture**: 284B total parameters with just 13B activated per forward pass, resulting in a compact 160GB download that runs on significantly less hardware than frontier models while maintaining strong performance.
* **1M-Token Context Window**: Shares the same 1-million-token context and 384K max output as V4-Pro, powered by the same CSA/HCA hybrid attention mechanism for efficient long-context inference.
* **Near-Pro Performance at 12.4x Lower Cost**: Scores 79.0% on SWE-bench Verified, only 1.6 percentage points behind V4-Pro's 80.6%, while costing 0.14/0.28 Credits per input/output token.
* **Flash-Max Reasoning Mode**: When given a larger thinking budget (384K+ context), V4-Flash-Max achieves comparable reasoning performance to V4-Pro, closing the gap on complex tasks.

## Best Use Cases

* **High-Volume API Workloads**: At 0.14 Credits per input token, Flash is ideal for applications that process large volumes of text where cost per query matters more than marginal accuracy gains.
* **Self-Hosted Deployments**: The 160GB model size and 13B active parameters make it feasible for on-premise or single-node GPU deployments, unlike larger frontier models.
* **Agentic Tool-Use Pipelines**: Strong tool-calling and coding capabilities paired with low latency make it well-suited for multi-step agent workflows where many LLM calls are chained together.

## Capabilities and Limitations

| Capability | Description |
| :----------------- | :-------------------------------------------------------------------------------------- |
| **Reasoning** | Competitive with Claude Sonnet 4.6 level intelligence (47 on Artificial Analysis Index) |
| **Coding** | 79.0% SWE-bench Verified; 64.4 average across coding benchmarks |
| **Multimodal** | Text-only; no image, audio, or video support |
| **Response Speed** | Optimized for high throughput with 13B active parameters and efficient attention |
| **Context Window** | 1,000,000 tokens |
| **Max Output** | 384,000 tokens |
| **Tool Use** | Function calling support; strong agentic task performance |
| **Multilingual** | Broad multilingual support; strongest in English and Chinese |

### Known Limitations

* Text-only, with no multimodal capabilities.
* Falls behind V4-Pro and frontier closed-source models on pure knowledge tasks and the most complex agentic workflows due to smaller parameter scale.
* May require Flash-Max mode (larger thinking budget) to match Pro-level reasoning, increasing latency and cost for complex tasks.

## Credits and Pricing

| Model | Input (Credits/Token) | Output (Credits/Token) | Notes |
| :-------------------- | --------------------: | ---------------------: | :-------------------------- |
| **DeepSeek V4 Flash** | `0.14` | `0.28` | 1M context, 384K max output |
43 changes: 43 additions & 0 deletions docs/llmservice/models/deepseek-v4-pro.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,43 @@
# DeepSeek V4 Pro

## Overview

DeepSeek-V4-Pro is DeepSeek's flagship open-source large language model, released on April 24, 2026 under the MIT License. Built on a Mixture-of-Experts (MoE) architecture with 1.6 trillion total parameters and 49 billion active parameters, it supports a 1-million-token context window and delivers frontier-level performance in coding, mathematics, and reasoning at a fraction of the cost of comparable closed-source models.

## Key Features

* **Massive Scale, Efficient Inference**: 1.6T total parameters with only 49B activated per forward pass via MoE, achieving frontier performance while keeping inference costs low, roughly 1/20th the price of Claude Opus 4.7.
* **1M-Token Context Window**: Supports up to 1 million input tokens and 384K maximum output tokens, enabled by a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) that reduces single-token inference FLOPs to 27% and KV cache to 10% compared to DeepSeek-V3.2.
* **Top-Tier Coding Performance**: Achieves 80.6% on SWE-bench Verified (within 0.2 points of Claude Opus 4.6), 67.9% on Terminal-Bench 2.0, 93.5% on LiveCodeBench, and a 3206 Codeforces rating.
* **Advanced Tool Use**: Scores 73.6 on MCPAtlas Public, supports up to 128 parallel function calls, and ships with pre-tuned adapters for Claude Code, OpenCode, OpenClaw, and CodeBuddy.

## Best Use Cases

* **Software Engineering Agents**: With 80.6% on SWE-bench Verified and strong tool-use capabilities, V4-Pro is well-suited for autonomous coding agents that need to navigate large codebases.
* **Long-Document Analysis**: The 1M-token context window enables processing of entire codebases, legal corpora, or research paper collections in a single pass.
* **Cost-Sensitive Frontier Workloads**: For teams that need near-frontier reasoning and coding at open-source economics, V4-Pro offers the best performance-per-dollar ratio in its class.

## Capabilities and Limitations

| Capability | Description |
| :----------------- | :------------------------------------------------------------------------ |
| **Reasoning** | Matches leading closed-source models on math and STEM; 95.2% on HMMT 2026 |
| **Coding** | 80.6% SWE-bench Verified, 93.5% LiveCodeBench, 3206 Codeforces rating |
| **Multimodal** | Text-only; multimodal capabilities are in development |
| **Response Speed** | Optimized for long-context efficiency via CSA/HCA hybrid attention |
| **Context Window** | 1,000,000 tokens |
| **Max Output** | 384,000 tokens |
| **Tool Use** | Function calling with up to 128 parallel calls; 73.6 on MCPAtlas Public |
| **Multilingual** | Broad multilingual support; strongest in English and Chinese |

### Known Limitations

* Text-only, with no image, audio, or video understanding or generation.
* Trails GPT-5.4 and Gemini 3.1 Pro on world knowledge benchmarks by a margin DeepSeek estimates at 3-6 months of development.
* May underperform on tasks requiring creative writing or highly nuanced reasoning compared to top closed-source models.

## Credits and Pricing

| Model | Input (Credits/Token) | Output (Credits/Token) | Notes |
| :------------------ | --------------------: | ---------------------: | :-------------------------- |
| **DeepSeek V4 Pro** | `1.74` | `3.48` | 1M context, 384K max output |
46 changes: 46 additions & 0 deletions docs/llmservice/models/gpt-5-5.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,46 @@
## Overview

GPT-5.5 is OpenAI's most capable model, released on April 23, 2026. Codenamed "Spud," it is the first fully retrained base model since GPT-4.5, and all prior GPT-5 releases were post-training iterations on the same foundation. GPT-5.5 features a natively omnimodal architecture, a 1-million-token context window, and leads on 14 benchmarks at launch.

## Key Features

* **Rebuilt Foundation**: First complete base model retraining since GPT-4.5, delivering fundamental capability improvements rather than incremental post-training gains across reasoning, coding, and knowledge tasks.
* **Natively Omnimodal**: A single parameter pool handles text, images, audio, and video, enabling seamless cross-modal reasoning without separate encoder modules.
* **60% Hallucination Reduction**: Achieves a 60% reduction in hallucination rate compared to GPT-5.4, and uses approximately 40% fewer output tokens on equivalent Codex tasks, improving both reliability and cost efficiency.
* **Frontier Benchmark Performance**: 88.7% on SWE-bench Verified, 92.4% on MMLU, 93.6% on GPQA Diamond, and 85.0% on ARC-AGI-2, an 11.7-point jump over GPT-5.4.

## Best Use Cases

* **Complex Software Engineering**: With 88.7% on SWE-bench Verified and strong tool-use integration, GPT-5.5 excels at end-to-end coding tasks including debugging, refactoring, and multi-file changes.
* **Multimodal Workflows**: The natively omnimodal architecture makes it uniquely suited for tasks that span text, images, audio, and video, such as analyzing meeting recordings, processing documents with figures, or building multimedia applications.
* **Research and Analysis**: Top scores on GPQA Diamond (93.6%) and MMLU (92.4%) make it the strongest choice for PhD-level science questions, complex reasoning, and knowledge-intensive research tasks.

## Capabilities and Limitations

| Capability | Description |
| :----------------- | :--------------------------------------------------------------------------------------- |
| **Reasoning** | 93.6% GPQA Diamond, 85.0% ARC-AGI-2, medium reasoning effort by default |
| **Coding** | 88.7% SWE-bench Verified; 40% fewer output tokens on Codex tasks vs GPT-5.4 |
| **Multimodal** | Natively omnimodal: text, image, audio, and video in a single architecture |
| **Response Speed** | Medium reasoning effort by default; configurable for latency-sensitive workloads |
| **Context Window** | 1,050,000 tokens (2x pricing for input beyond 272K tokens) |
| **Max Output** | 128,000 tokens |
| **Tool Use** | Full function calling, tool search, hosted tools, prompt caching, and compaction support |
| **Multilingual** | Broad multilingual support across major languages |

### Known Limitations

* Significantly more expensive than open-source alternatives, roughly 50x the cost of DeepSeek V4 Pro for input tokens.
* Lost the harder SWE-Bench Pro benchmark to Claude Opus 4.7 despite winning the standard SWE-Bench Verified headline.
* 2x input pricing for prompts exceeding 272K tokens increases costs substantially for long-context workloads.

## Credits and Pricing

| Model | Input (Credits/Token) | Output (Credits/Token) | Notes |
| :---------- | --------------------: | ---------------------: | :---------------------------------------- |
| **GPT-5.5** | `5.00` | `30.00` | 2x input / 1.5x output beyond 272K tokens |

## References

[1] [OpenAI GPT-5.5 Announcement](https://openai.com/index/gpt-5-5)
[2] [OpenAI API Pricing](https://openai.com/api/pricing)
3 changes: 3 additions & 0 deletions docs/llmservice/pricing-and-usage.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,10 @@ The platform utilizes a unified credit system to measure and settle usage for al
| Kimi K2.5 | 0.23 | 3.00 | - |
| GLM-5 | 0.30 | 2.55 | - |
| DeepSeek V3.2 | 0.27 | 0.42 | - |
| DeepSeek V4 Flash | 0.14 | 0.28 | - |
| DeepSeek V4 Pro | 1.74 | 3.48 | - |
| GPT-5.4 | 2.50 | 15.00 | 10,000 |
| GPT-5.5 | 5.00 | 30.00 | - |
| GPT-5.4 Pro | 30.00 | 180.00 | - |
| GPT-5.2 | 1.75 | 14.00 | 10,000 |
| GPT-5.4 Mini | 0.75 | 4.50 | 10,000 |
Expand Down
43 changes: 43 additions & 0 deletions ...s/docusaurus-plugin-content-docs/current/llmservice/models/deepseek-v4-flash.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,43 @@
# DeepSeek V4 Flash

## 概述

DeepSeek-V4-Flash 是 DeepSeek 于 2026 年 4 月 24 日与 V4-Pro 同步发布的高效率开源大语言模型，采用 MIT License。该模型总参数量为 284B，但每次前向仅激活 13B 参数，以仅为 V4-Pro 约 1/12.4 的成本提供接近旗舰模型的性能，是当前极具性价比的模型之一。

## 核心特性

* **超高效率架构**：总参数量 284B，每次前向仅激活 13B 参数，模型下载体积约 160GB，相比前沿模型对硬件要求更低，同时保持出色性能。
* **100 万 Token 上下文窗口**：与 V4-Pro 一样支持 100 万上下文和 384K 最大输出，基于相同的 CSA/HCA 混合注意力机制，具备高效的长上下文推理能力。
* **接近 Pro 的性能与 12.4 倍更低成本**：在 SWE-bench Verified 上达到 79.0%，仅比 V4-Pro 的 80.6% 低 1.6 个百分点，而输入/输出价格仅为 0.14 / 0.28 Credits。
* **Flash-Max 推理模式**：在提供更大的思考预算（384K+ 上下文）时，V4-Flash-Max 可在复杂任务上逼近 V4-Pro 的推理能力。

## 适用场景

* **高并发 API 场景**：以每输入 token 仅 0.14 Credits 的成本，非常适合文本量大、对单次调用成本敏感的应用。
* **自托管部署**：160GB 模型体积和 13B 激活参数使其更适合本地部署或单节点 GPU 场景，不像更大的前沿模型那样依赖重型基础设施。
* **Agent 工具调用链路**：强工具调用和编程能力，加上更低延迟，使其非常适合多步 Agent 工作流。

## 能力与限制

| 能力维度 | 说明 |
| :--- | :--- |
| **推理能力** | 具备接近 Claude Sonnet 4.6 水平的智能表现（Artificial Analysis Index 得分 47） |
| **编程能力** | SWE-bench Verified 79.0%；编程类基准平均 64.4 |
| **多模态能力** | 当前仅支持文本，不支持图像、音频或视频 |
| **响应速度** | 13B 激活参数配合高效注意力机制，适合高吞吐场景 |
| **上下文窗口** | 1,000,000 tokens |
| **最大输出** | 384,000 tokens |
| **工具调用** | 支持函数调用，具备较强 Agent 任务执行能力 |
| **多语言能力** | 广泛支持多语言，其中英文表现最强 |

### 已知限制

* 当前仅支持文本，不具备多模态能力。
* 由于参数规模更小，在纯知识型任务和最复杂的 Agent 工作流上仍弱于 V4-Pro 与顶级闭源模型。
* 若要接近 Pro 级推理效果，可能需要启用 Flash-Max 模式并提供更大思考预算，这会带来更高延迟和成本。

## 积分与定价

| 模型名称 | 输入 (Credits/Token) | 输出 (Credits/Token) | 备注 |
| :--- | --------------------: | -------------------: | :--- |
| **DeepSeek V4 Flash** | `0.14` | `0.28` | 100 万上下文，384K 最大输出 |
43 changes: 43 additions & 0 deletions ...ans/docusaurus-plugin-content-docs/current/llmservice/models/deepseek-v4-pro.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,43 @@
# DeepSeek V4 Pro

## 概述

DeepSeek-V4-Pro 是 DeepSeek 于 2026 年 4 月 24 日基于 MIT License 发布的旗舰级开源大语言模型。该模型采用 Mixture-of-Experts（MoE）架构，总参数量达到 1.6T，单次前向激活参数约为 49B，支持 100 万 token 上下文窗口，并以远低于同类闭源模型的成本，提供接近前沿水平的编程、数学与推理能力。

## 核心特性

* **超大规模与高效推理**：总参数量 1.6T，但每次推理仅激活 49B 参数，在保持前沿性能的同时显著降低推理成本，整体价格约为 Claude Opus 4.7 的 1/20。
* **100 万 Token 上下文窗口**：支持最高 100 万输入 token 与 384K 最大输出 token。通过结合 Compressed Sparse Attention（CSA）与 Heavily Compressed Attention（HCA）的混合注意力机制，相比 DeepSeek-V3.2 可将单 token 推理 FLOPs 降至 27%，KV Cache 降至 10%。
* **顶级编程能力**：在 SWE-bench Verified 上达到 80.6%，与 Claude Opus 4.6 仅差 0.2 分；同时在 Terminal-Bench 2.0 上达到 67.9%，在 LiveCodeBench 上达到 93.5%，Codeforces 评分达到 3206。
* **高级工具调用能力**：在 MCPAtlas Public 上得分 73.6，支持最多 128 个并行函数调用，并针对 Claude Code、OpenCode、OpenClaw 与 CodeBuddy 提供了预调优适配器。

## 适用场景

* **软件工程 Agent**：凭借 80.6% 的 SWE-bench Verified 成绩和强工具调用能力，V4-Pro 非常适合需要遍历大型代码库的自主编程 Agent。
* **长文档分析**：100 万 token 上下文窗口使其能够一次性处理整套代码库、法律文档集合或科研论文集。
* **成本敏感的前沿任务**：对于既需要接近前沿的推理与编程能力、又关注成本效率的团队，V4-Pro 提供了极具竞争力的性能价格比。

## 能力与限制

| 能力维度 | 说明 |
| :--- | :--- |
| **推理能力** | 在数学与 STEM 任务上可与顶级闭源模型对标；HMMT 2026 达到 95.2% |
| **编程能力** | SWE-bench Verified 80.6%，LiveCodeBench 93.5%，Codeforces 3206 |
| **多模态能力** | 当前仅支持文本，多模态能力仍在开发中 |
| **响应速度** | 通过 CSA/HCA 混合注意力机制优化长上下文推理效率 |
| **上下文窗口** | 1,000,000 tokens |
| **最大输出** | 384,000 tokens |
| **工具调用** | 支持最多 128 个并行函数调用；MCPAtlas Public 得分 73.6 |
| **多语言能力** | 具备广泛多语言支持，其中英文表现最强 |

### 已知限制

* 当前仅支持文本，不支持图像、音频或视频的理解与生成。
* 在世界知识类基准上，较 GPT-5.4 与 Gemini 3.1 Pro 仍有差距，DeepSeek 估计约相差 3-6 个月研发进度。
* 在创意写作或高度细腻推理任务上，可能弱于顶级闭源模型。

## 积分与定价

| 模型名称 | 输入 (Credits/Token) | 输出 (Credits/Token) | 备注 |
| :--- | --------------------: | -------------------: | :--- |
| **DeepSeek V4 Pro** | `1.74` | `3.48` | 100 万上下文，384K 最大输出 |
Loading
Loading