ruflo 中 Flow Nexus Neural 技能解析:MCP 驱动的分布式神经网络训练、推理与模型市场
本文基于 ruflo 仓库中的 Flow Nexus Neural 技能定义(SKILL.md)展开,完整讲解如何通过 Flow Nexus MCP 工具集完成单节点训练、模型推理、模板市场部署、E2B 沙箱分布式训练集群搭建与联邦学习,并补充仓库内配套的 Agent 定义、命令参考与认证流程作为源码级佐证。读完后你可以独立编写神经网络训练配置(feedforward / LSTM / Transformer / GAN / autoencoder)、编排多节点训练集群,并掌握模板发布、性能基准测试与故障排查的完整实操链路。
技能定位与前置条件
该技能在 frontmatter 中声明了关键元数据:名称 flow-nexus-neural、类别 ai-ml、版本 1.0.0,并带有 neural-networks、distributed-training、flow-nexus、e2b-sandboxes 等标签。两个约束字段值得注意:
requires_auth: true—— 所有功能(尤其是分布式训练)依赖 Flow Nexus 账户认证;mcp_server: flow-nexus—— 技能的全部能力由 Flow Nexus MCP 服务器提供的工具实现,本地不执行训练,而是把任务编排到云端 E2B 沙箱环境中。
环境准备
接入分两步:先注册 MCP 服务器,再完成账户注册与登录:
# Add Flow Nexus MCP server
claude mcp add flow-nexus npx flow-nexus@latest mcp start
# Register and login
npx flow-nexus@latest register
npx flow-nexus@latest login
仓库中的 认证命令参考 进一步给出了 MCP 工具侧的认证接口:mcp__flow-nexus__user_register(邮箱 + 密码注册)、mcp__flow-nexus__user_login(登录)、mcp__flow-nexus__auth_status({ detailed: true })(检查认证状态)、mcp__flow-nexus__user_logout(登出),以及 user_reset_password / user_update_password 的密码重置流程。完整路径是:CLI 注册登录 → MCP 认证状态可用 → 调用 neural 系列工具。
能力一:单节点神经网络训练(neural_train)
neural_train 是技能的入口工具,支持五种架构与五档资源规格。
可用架构(config.architecture.type):
| 架构 | 适用场景 |
|---|---|
feedforward |
标准全连接网络,分类/回归/简单模式识别 |
lstm |
序列建模,时间序列预测 |
gan |
生成对抗网络,图像合成等生成任务 |
autoencoder |
降维、异常检测 |
transformer |
基于注意力的 NLP / 多模态模型 |
训练档位(tier): nano(最小资源,快速验证)→ mini(小模型)→ small(标准)→ medium(复杂模型)→ large(大规模训练)。
示例 1:自定义前馈分类器
完整配置展示了层级的堆叠方式——dense 层与 dropout 层交替,末层使用 softmax 输出 10 类:
mcp__flow-nexus__neural_train({
config: {
architecture: {
type: "feedforward",
layers: [
{ type: "dense", units: 256, activation: "relu" },
{ type: "dropout", rate: 0.3 },
{ type: "dense", units: 128, activation: "relu" },
{ type: "dropout", rate: 0.2 },
{ type: "dense", units: 64, activation: "relu" },
{ type: "dense", units: 10, activation: "softmax" }
]
},
training: {
epochs: 100,
batch_size: 32,
learning_rate: 0.001,
optimizer: "adam"
},
divergent: {
enabled: true,
pattern: "lateral", // quantum, chaotic, associative, evolutionary
factor: 0.5
}
},
tier: "small",
user_id: "your_user_id"
})
注意 divergent 块:这是该技能特有的"发散式训练"选项,pattern 可取 lateral(侧向)、quantum、chaotic、associative、evolutionary 五类探索模式,factor 控制发散强度,可用于跳出局部最优的探索性训练。
示例 2:LSTM 时序预测
mcp__flow-nexus__neural_train({
config: {
architecture: {
type: "lstm",
layers: [
{ type: "lstm", units: 128, return_sequences: true },
{ type: "dropout", rate: 0.2 },
{ type: "lstm", units: 64 },
{ type: "dense", units: 1, activation: "linear" }
]
},
training: {
epochs: 150,
batch_size: 64,
learning_rate: 0.01,
optimizer: "adam"
}
},
tier: "medium"
})
关键点:第一层 LSTM 设置 return_sequences: true 以把序列输出传递给第二层 LSTM,构成堆叠结构;末层 dense units: 1 + linear 是典型的回归型单值预测输出。
示例 3:Transformer 文本模型
mcp__flow-nexus__neural_train({
config: {
architecture: {
type: "transformer",
layers: [
{ type: "embedding", vocab_size: 10000, embedding_dim: 512 },
{ type: "transformer_encoder", num_heads: 8, ff_dim: 2048 },
{ type: "global_average_pooling" },
{ type: "dense", units: 128, activation: "relu" },
{ type: "dense", units: 2, activation: "softmax" }
]
},
training: {
epochs: 50,
batch_size: 16,
learning_rate: 0.0001,
optimizer: "adam"
}
},
tier: "large"
})
这是"嵌入 → 单头组编码器(8 头、前馈维度 2048)→ 全局平均池化 → 双类分类头"的二分类文本模型。由于 Transformer 显存与计算开销大,示例直接选用 large 档位;相应地,学习率也降到 0.0001。
仓库中的 命令参考 还给出了两种"简写配置":省略 layers 只写 { architecture: { type: "cnn" } } 或 { type: "lstm" },说明 MCP 层支持按架构类型采用默认层栈,适合快速实验。
能力二:模型推理(neural_predict)
训练完成后通过 model_id 执行批量预测,输入为二维数组(多条样本):
mcp__flow-nexus__neural_predict({
model_id: "model_abc123",
input: [
[0.5, 0.3, 0.2, 0.1],
[0.8, 0.1, 0.05, 0.05],
[0.2, 0.6, 0.15, 0.05]
],
user_id: "your_user_id"
})
响应结构包含逐样本预测、推理耗时与模型版本:
{
"predictions": [
[0.12, 0.85, 0.03],
[0.89, 0.08, 0.03],
[0.05, 0.92, 0.03]
],
"inference_time_ms": 45,
"model_version": "1.0.0"
}
在 NLP 场景中 input 可以直接是字符串数组(见下文"情感分析"用例),即同一工具兼容数值张量与文本输入,由模型侧做预处理。
能力三:模板市场(Template Marketplace)
列出模板
neural_list_templates 支持按类别、付费档位、关键词过滤:
mcp__flow-nexus__neural_list_templates({
category: "classification", // timeseries, regression, nlp, vision, anomaly, generative
tier: "free", // or "paid"
search: "sentiment",
limit: 20
})
category 可选值:classification、timeseries、regression、nlp、vision、anomaly、generative;tier 取 free 或 paid。响应示例:
{
"templates": [
{
"id": "sentiment-analysis-v2",
"name": "Sentiment Analysis Classifier",
"description": "Pre-trained BERT model for sentiment analysis",
"category": "nlp",
"accuracy": 0.94,
"downloads": 1523,
"tier": "free"
},
{
"id": "image-classifier-resnet",
"name": "ResNet Image Classifier",
"description": "ResNet-50 for image classification",
"category": "vision",
"accuracy": 0.96,
"downloads": 2341,
"tier": "paid"
}
]
}
部署模板
部署时可叠加 custom_config 覆盖训练参数(如微调轮数与学习率):
mcp__flow-nexus__neural_deploy_template({
template_id: "sentiment-analysis-v2",
custom_config: {
training: {
epochs: 50,
learning_rate: 0.0001
}
},
user_id: "your_user_id"
})
能力四:分布式训练集群
这是该技能区别于普通单机训练的核心:把训练任务拆到多个 E2B 沙箱节点上执行。完整生命周期为:初始化集群 → 部署节点 → 连接拓扑 → 启动分布式训练 → 监控 → 分布式推理 → 终止。
1. 初始化集群(neural_cluster_init)
mcp__flow-nexus__neural_cluster_init({
name: "large-model-cluster",
architecture: "transformer", // transformer, cnn, rnn, gnn, hybrid
topology: "mesh", // mesh, ring, star, hierarchical
consensus: "proof-of-learning", // byzantine, raft, gossip
daaEnabled: true, // Decentralized Autonomous Agents
wasmOptimization: true
})
参数语义:architecture 覆盖 transformer / cnn / rnn / gnn / hybrid 五类;topology 支持 mesh(全连接网状)、ring(环形)、star(星型)、hierarchical(分层)四种通信拓扑;consensus 指定节点间共识协议,proof-of-learning 之外还有 byzantine、raft、gossip 可选;daaEnabled 开启去中心化自治代理模式,wasmOptimization 启用 WASM 优化以降低节点开销。
{
"cluster_id": "cluster_xyz789",
"name": "large-model-cluster",
"status": "initializing",
"topology": "mesh",
"max_nodes": 100,
"created_at": "2025-10-19T10:30:00Z"
}
2. 部署节点(neural_node_deploy)
一个典型集群由三类角色构成——参数服务器、工作节点、聚合器:
// Deploy parameter server
mcp__flow-nexus__neural_node_deploy({
cluster_id: "cluster_xyz789",
node_type: "parameter_server",
model: "large",
template: "nodejs",
capabilities: ["parameter_management", "gradient_aggregation"],
autonomy: 0.8
})
// Deploy worker nodes
mcp__flow-nexus__neural_node_deploy({
cluster_id: "cluster_xyz789",
node_type: "worker",
model: "xl",
role: "worker",
capabilities: ["training", "inference"],
layers: [
{ type: "transformer_encoder", num_heads: 16 },
{ type: "feed_forward", units: 4096 }
],
autonomy: 0.9
})
// Deploy aggregator
mcp__flow-nexus__neural_node_deploy({
cluster_id: "cluster_xyz789",
node_type: "aggregator",
model: "large",
capabilities: ["gradient_aggregation", "model_synchronization"]
})
从参数设计看:node_type 取 parameter_server / worker / aggregator;model 是节点计算规格(示例中出现 large、xl);autonomy(0~1)控制节点的自治程度,与集群级 daaEnabled 呼应;capabilities 声明节点职能;worker 节点还可直接内联 layers 定义其承载的模型分片结构。
3. 连接拓扑与启动训练
// 连接拓扑(覆盖初始化时的默认值)
mcp__flow-nexus__neural_cluster_connect({
cluster_id: "cluster_xyz789",
topology: "mesh"
})
// 启动分布式训练
mcp__flow-nexus__neural_train_distributed({
cluster_id: "cluster_xyz789",
dataset: "imagenet", // or custom dataset identifier
epochs: 100,
batch_size: 128,
learning_rate: 0.001,
optimizer: "adam", // sgd, rmsprop, adagrad
federated: true // Enable federated learning
})
dataset 可传内置数据集标识(如 imagenet)或自定义数据集标识;optimizer 支持 adam、sgd、rmsprop、adagrad。
联邦学习变体:数据保留在各本地节点,仅聚合参数,附加 aggregation_rounds 与 min_nodes_per_round 两个控制参数:
mcp__flow-nexus__neural_train_distributed({
cluster_id: "cluster_xyz789",
dataset: "medical_images_distributed",
epochs: 200,
batch_size: 64,
learning_rate: 0.0001,
optimizer: "adam",
federated: true, // Data stays on local nodes
aggregation_rounds: 50,
min_nodes_per_round: 5
})
4. 监控与终止
mcp__flow-nexus__neural_cluster_status({
cluster_id: "cluster_xyz789"
})
{
"cluster_id": "cluster_xyz789",
"status": "training",
"nodes": [
{
"node_id": "node_001",
"type": "parameter_server",
"status": "active",
"cpu_usage": 0.75,
"memory_usage": 0.82
},
{
"node_id": "node_002",
"type": "worker",
"status": "active",
"training_progress": 0.45
}
],
"training_metrics": {
"current_epoch": 45,
"total_epochs": 100,
"loss": 0.234,
"accuracy": 0.891
}
}
状态响应同时给出集群级 training_metrics(epoch / loss / accuracy)与节点级资源占用(cpu / memory / 训练进度),适合写入定时监控循环。集群任务结束时调用 mcp__flow-nexus__neural_cluster_terminate({ cluster_id: "cluster_xyz789" }) 释放沙箱资源。
5. 分布式推理
多节点结果可经聚合策略合并:
mcp__flow-nexus__neural_predict_distributed({
cluster_id: "cluster_xyz789",
input_data: JSON.stringify([
[0.1, 0.2, 0.3],
[0.4, 0.5, 0.6]
]),
aggregation: "ensemble" // mean, majority, weighted, ensemble
})
注意 input_data 需要 JSON.stringify 序列化为字符串传入;aggregation 支持 mean(均值)、majority(多数表决)、weighted(加权)、ensemble(集成)四种合并方式。
能力五:模型管理
列出模型
mcp__flow-nexus__neural_list_models({
user_id: "your_user_id",
include_public: true
})
{
"models": [
{
"model_id": "model_abc123",
"name": "Custom Classifier v1",
"architecture": "feedforward",
"accuracy": 0.92,
"created_at": "2025-10-15T14:20:00Z",
"status": "trained"
},
{
"model_id": "model_def456",
"name": "LSTM Forecaster",
"architecture": "lstm",
"mse": 0.0045,
"created_at": "2025-10-18T09:15:00Z",
"status": "training"
}
]
}
模型条目包含架构类型、评价指标(分类任务为 accuracy,回归任务为 mse)与状态(training / trained)。
训练状态查询
mcp__flow-nexus__neural_training_status({
job_id: "job_training_xyz"
})
{
"job_id": "job_training_xyz",
"status": "training",
"progress": 0.67,
"current_epoch": 67,
"total_epochs": 100,
"current_loss": 0.234,
"estimated_completion": "2025-10-19T12:45:00Z"
}
job_id 来自 neural_train 的返回,响应提供进度比、当前 epoch、损失与预计完成时间。
性能基准测试与验证工作流
// 基准测试
mcp__flow-nexus__neural_performance_benchmark({
model_id: "model_abc123",
benchmark_type: "comprehensive" // inference, throughput, memory, comprehensive
})
{
"model_id": "model_abc123",
"benchmarks": {
"inference_latency_ms": 12.5,
"throughput_qps": 8000,
"memory_usage_mb": 245,
"gpu_utilization": 0.78,
"accuracy": 0.92,
"f1_score": 0.89
},
"timestamp": "2025-10-19T11:00:00Z"
}
benchmark_type 可选 inference(延迟)、throughput(吞吐)、memory(内存)、comprehensive(全项)。生产部署前应使用 comprehensive 全面评估。
// 创建验证工作流
mcp__flow-nexus__neural_validation_workflow({
model_id: "model_abc123",
user_id: "your_user_id",
validation_type: "comprehensive" // performance, accuracy, robustness, comprehensive
})
validation_type 覆盖性能、精度、鲁棒性三维验证。
能力六:模型发布与评分
把训练好的模型发布为市场模板,price: 0 表示免费,否则为点数价格:
mcp__flow-nexus__neural_publish_template({
model_id: "model_abc123",
name: "High-Accuracy Sentiment Classifier",
description: "Fine-tuned BERT model for sentiment analysis with 94% accuracy",
category: "nlp",
price: 0, // 0 for free, or credits amount
user_id: "your_user_id"
})
// 对模板评分
mcp__flow-nexus__neural_rate_template({
template_id: "sentiment-analysis-v2",
rating: 5,
review: "Excellent model! Achieved 95% accuracy on my dataset.",
user_id: "your_user_id"
})
典型组合用例
图像分类(CNN + 分层拓扑集群)
// Initialize cluster for large-scale image training
const cluster = await mcp__flow-nexus__neural_cluster_init({
name: "image-classification-cluster",
architecture: "cnn",
topology: "hierarchical",
wasmOptimization: true
})
// Deploy worker nodes
await mcp__flow-nexus__neural_node_deploy({
cluster_id: cluster.cluster_id,
node_type: "worker",
model: "large",
capabilities: ["training", "data_augmentation"]
})
// Start training
await mcp__flow-nexus__neural_train_distributed({
cluster_id: cluster.cluster_id,
dataset: "custom_images",
epochs: 100,
batch_size: 64,
learning_rate: 0.001,
optimizer: "adam"
})
NLP 情感分析(模板 + 文本推理)
// Use pre-built template
const deployment = await mcp__flow-nexus__neural_deploy_template({
template_id: "sentiment-analysis-v2",
custom_config: {
training: {
epochs: 30,
batch_size: 16
}
}
})
// Run inference
const result = await mcp__flow-nexus__neural_predict({
model_id: deployment.model_id,
input: ["This product is amazing!", "Terrible experience."]
})
时间序列预测(LSTM + 进度监控)
// Train LSTM model
const training = await mcp__flow-nexus__neural_train({
config: {
architecture: {
type: "lstm",
layers: [
{ type: "lstm", units: 128, return_sequences: true },
{ type: "dropout", rate: 0.2 },
{ type: "lstm", units: 64 },
{ type: "dense", units: 1 }
]
},
training: {
epochs: 150,
batch_size: 64,
learning_rate: 0.01,
optimizer: "adam"
}
},
tier: "medium"
})
// Monitor progress
const status = await mcp__flow-nexus__neural_training_status({
job_id: training.job_id
})
隐私敏感场景的联邦学习
// Initialize federated cluster
const cluster = await mcp__flow-nexus__neural_cluster_init({
name: "federated-medical-cluster",
architecture: "transformer",
topology: "mesh",
consensus: "proof-of-learning",
daaEnabled: true
})
// Deploy nodes across different locations
for (let i = 0; i < 5; i++) {
await mcp__flow-nexus__neural_node_deploy({
cluster_id: cluster.cluster_id,
node_type: "worker",
model: "large",
autonomy: 0.9
})
}
// Train with federated learning (data never leaves nodes)
await mcp__flow-nexus__neural_train_distributed({
cluster_id: cluster.cluster_id,
dataset: "medical_records_distributed",
epochs: 200,
federated: true,
aggregation_rounds: 100
})
该模式适用于医疗数据等"数据不可出域"场景:federated: true 保证数据保留在本地节点,集群仅交换聚合参数。
架构模式速查表
文档给出了五种架构的最小配置骨架,可直接作为 config.architecture 的模板:
Feedforward(分类、回归、简单模式识别):
{
type: "feedforward",
layers: [
{ type: "dense", units: 256, activation: "relu" },
{ type: "dropout", rate: 0.3 },
{ type: "dense", units: 128, activation: "relu" },
{ type: "dense", units: 10, activation: "softmax" }
]
}
LSTM(时间序列、序列、预测):
{
type: "lstm",
layers: [
{ type: "lstm", units: 128, return_sequences: true },
{ type: "lstm", units: 64 },
{ type: "dense", units: 1 }
]
}
Transformer(NLP、注意力机制、大规模文本):
{
type: "transformer",
layers: [
{ type: "embedding", vocab_size: 10000, embedding_dim: 512 },
{ type: "transformer_encoder", num_heads: 8, ff_dim: 2048 },
{ type: "global_average_pooling" },
{ type: "dense", units: 2, activation: "softmax" }
]
}
GAN(生成任务、图像合成):
{
type: "gan",
generator_layers: [...],
discriminator_layers: [...]
}
GAN 配置使用 generator_layers 与 discriminator_layers 双字段而非统一的 layers,分别描述生成器与判别器网络。
Autoencoder(降维、异常检测):
{
type: "autoencoder",
encoder_layers: [
{ type: "dense", units: 128, activation: "relu" },
{ type: "dense", units: 64, activation: "relu" }
],
decoder_layers: [
{ type: "dense", units: 128, activation: "relu" },
{ type: "dense", units: input_dim, activation: "sigmoid" }
]
}
编码器逐层降维、解码器镜像升维,末层激活通常为 sigmoid(对 0~1 归一化输入)。
仓库源码佐证:技能在 ruflo 中的配套资产
该 SKILL 并非孤立文件,仓库内存在多层配套定义,可以从源码结构确认其协作关系:
- Agent 定义:plugin/agents/flow-nexus/neural-network.md 定义了
flow-nexus-neural智能体角色,其职责与技能文档一一对应——架构设计、多沙箱分布式训练编排、模型生命周期管理、参数与资源优化、版本化/验证/基准测试,以及联邦学习与共识协议实现。Agent 中还给出六步工作流(问题分析 → 架构设计 → 资源规划 → 训练编排 → 模型验证 → 部署管理)和六类专项架构(Feedforward、LSTM/RNN、Transformer、CNN、GAN、Autoencoder),以及高阶能力清单:跨 E2B 沙箱分布式训练、隐私保护联邦学习、模型压缩、迁移学习、集成方法与模型漂移监测。 - 命令参考:plugin/commands/flow-nexus/neural-network.md 提供了同一组
mcp__flow-nexus__neural_*工具的速查版调用示例,其中额外演示了 CNN 与 LSTM 的简写配置形态,可视为该技能工具面的官方摘要。 - 认证命令:plugin/commands/flow-nexus/login-registration.md 对应 SKILL 中
requires_auth: true的前置要求,提供user_register/user_login/auth_status/user_logout/ 密码重置等 MCP 工具用法。 - 本地训练技能互补:仓库另有一个 neural-training 技能(frontmatter 声明用于 SONA / MoE / EWC++ 模式的本地模式训练),它通过
npx claude-flow neural train|status|patterns|predict命令操作。从两者分工看,flow-nexus-neural面向云端 E2B 沙箱的重型训练与集群编排,neural-training面向 Agent 内部的轻量模式学习,属于不同执行面,不宜混用。 - Swarm 编排关联:flow-nexus-swarm 技能 展示了同一 Flow Nexus 平台的 swarm/workflow 面(
swarm_init等工具,支持 mesh/ring/star/hierarchical 拓扑),与 neural 集群的topology取值体系一致,说明集群通信拓扑是平台级的统一抽象。
最佳实践与故障排查
八条最佳实践
- Start Small:实验阶段先用
nano或mini档位验证配置; - Use Templates:常见任务优先用市场模板起步;
- Monitor Training:定期查询状态(
neural_training_status/neural_cluster_status)及早发现问题; - Benchmark Models:生产部署前必须跑
comprehensive基准; - Distributed Training:大模型(文档建议 >1B 参数)使用集群训练;
- Federated Learning:隐私敏感数据启用联邦模式;
- Version Models:将验证通过的模型发布为模板复用;
- Validate Thoroughly:部署前走完整验证工作流。
训练停滞(Training Stalled)
先看集群状态定位卡住的节点,必要时终止重建:
// Check cluster status
const status = await mcp__flow-nexus__neural_cluster_status({
cluster_id: "cluster_id"
})
// Terminate and restart if needed
await mcp__flow-nexus__neural_cluster_terminate({
cluster_id: "cluster_id"
})
精度偏低(Low Accuracy)
按顺序排查:增加 epochs → 调整学习率 → 增加 dropout 正则化 → 更换 optimizer → 启用数据增强。
内存不足(Out of Memory)
减小 batch_size → 降低模型档位(tier)→ 启用梯度累积 → 切换到分布式训练分摊显存压力。
适用前提与边界说明
- 技能整体依赖外部 Flow Nexus 服务:
mcp_server: flow-nexus提供的 MCP 工具与 E2B 云端沙箱,本地仓库不包含训练运行时,所有mcp__flow-nexus__neural_*调用都在平台侧执行; requires_auth: true意味着未注册/未登录账户无法使用分布式训练等核心能力;- 文档中"Resources"一节指向的 Flow Nexus 文档站、模板市场与 API 参考为外部平台地址(原文以占位形式书写),本文按规范不输出外部链接,相关入口以 MCP 工具响应与
npx flow-nexus@latestCLI 为准。
相关文档
- 技能本体:SKILL.md(含插件镜像 plugin/skills/flow-nexus-neural/SKILL.md)
- 配套 Agent 角色定义:plugin/agents/flow-nexus/neural-network.md
- 命令速查:plugin/commands/flow-nexus/neural-network.md
- 认证与账户管理:plugin/commands/flow-nexus/login-registration.md
- 本地模式训练技能(互补):.agents/skills/neural-training/SKILL.md
- 同平台 swarm 编排:.agents/skills/flow-nexus-swarm/SKILL.md
atomcodeClaude Code 的开源替代方案。连接任意大模型,编辑代码,运行命令,自动验证 — 全自动执行。用 Rust 构建,极致性能。 | An open-source alternative to Claude Code. Connect any LLM, edit code, run commands, and verify changes — autonomously. Built in Rust for speed. Get StartedRust0627
Hy4-previewHy4 preview 是由腾讯混元团队研发的新一代混合专家(MoE)旗舰模型。模型总参数量 770B,每个 token 激活 49B,主干共包含78层,第一层采用标准 FFN,其余 77 层均为 MoE 结构,每层包含 256 个路由专家与 1 个共享专家,每个 token 激活 top-8 路由专家及共享专家。主干之外原生内置 1 层 MTP(总参数量 10B,激活 0.7B)以支持投机解码。Python00
GLM-5.3GLM-5.3 与 GLM-5.2 使用相同的基座模型——所有提升均来自后训练。与 GLM-5.2 相比,它在复杂编程和长程任务上的表现显著提升。Jinja00
GLM-5.3-FlashGLM-5.3-Flash (320B-A18B),是GLM-5系列的首个原生多模态模型。320B总参数,能力超过GLM-5.2Jinja00
Spark-X2.5-4BSpark-X2.5-4B 旨在让强大的 AI 更实用、更高效、更易获得。在广泛日常任务中表现强劲,涵盖对话、写作、翻译、推理、编码、工具调用以及智能体工作流,并在同等规模的开源模型中取得领先成绩。Spark-X2.5 将面向效率的架构与最高 1M tokens 的原生上下文窗口相结合,并支持 200 多种语言。Python00
Spark-X2.5-1.7BSpark-X2.5-1.7B 旨在让强大的 AI 更加实用、高效且易于获取。这些模型在广泛的日常任务中表现出色,涵盖对话、写作、翻译、推理、编程、工具调用和智能体工作流,并在同等规模的开源模型中取得领先结果。Spark-X2.5 将面向效率的架构与最高 1M tokens 的原生上下文窗口相结合,并支持 200 多种语言。Python00