AI News AI资讯 3h ago Updated 1h ago 更新于 1小时前 49

IBM Releases Granite 4.2: Bringing Native Reasoning and Agentic RL to Open Enterprise Models IBM发布Granite 4.2:将原生推理和Agentic RL引入开源企业模型

IBM released Granite 4.2, a family of open reasoning LLMs (3B, 8B, 30B) built from scratch on ~15T tokens, with native chain-of-thought reasoning and a thinking/non-thinking/low-effort mode switch Agentic RL (SWE, Terminal, Search) is applied only to the 8B and 30B models via a multi-stage asynchronous GRPO pipeline, creating a significant capability gap between the 3B and larger variants All models ship under Apache 2.0 with full commercial deployment rights, targeting everything from laptop in IBM发布Granite 4.2开源推理语言模型家族,包含3B、8B和30B三种参数规模,全面采用Apache 2.0许可证 模型架构为解码器密集Transformer,预训练约15万亿token,支持显式思维链推理及thinking/non-thinking/低开销模式切换 8B和30B版本引入Agentic RL训练阶段,在真实沙盒环境中学习代码编辑、终端操作和网页搜索能力 同步发布Granite Speech 5.0 Turbo CTC语音模型(470M参数),在H200上实现约12,600 RTFx吞吐量,较现有速度领先者提升约一倍 30B模型在SWE-Bench Verified达到

72
Hot 热度
70
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • IBM released Granite 4.2, a family of open reasoning LLMs (3B, 8B, 30B) built from scratch on ~15T tokens, with native chain-of-thought reasoning and a thinking/non-thinking/low-effort mode switch
  • Agentic RL (SWE, Terminal, Search) is applied only to the 8B and 30B models via a multi-stage asynchronous GRPO pipeline, creating a significant capability gap between the 3B and larger variants
  • All models ship under Apache 2.0 with full commercial deployment rights, targeting everything from laptop inference (3B via Ollama/LM Studio) to enterprise GPU clusters (30B via vLLM)
  • The 30B model achieves 57.00 on SWE-Bench Verified and 29.24 on Terminal-Bench 2.1, while the 3B already reaches 78.33 on AIME25
  • IBM simultaneously released Granite Speech 5.0 Turbo CTC, a 470M-parameter ASR model with no LLM backbone, delivering ~12,600 RTFx throughput on a single H200

Why It Matters

Granite 4.2 represents a significant shift in the open-weight reasoning model landscape by combining explicit chain-of-thought capabilities with agentic reinforcement learning at scale, all under a permissive Apache 2.0 license. For AI practitioners, it provides a rare deployable alternative to closed reasoning models, with size-appropriate options ranging from edge devices to data-center GPUs, and the agentic RL training on real sandboxed environments directly addresses the growing demand for autonomous coding and tool-use agents in production.

Technical Details

  • Architecture: Decoder-only dense transformer (not MoE or hybrid), using Grouped Query Attention with 8 KV heads, RoPE (θ = 10,000,000), SwiGLU MLPs, RMSNorm (ε = 1e-5), untied input/output embeddings, and bfloat16 precision. The 3B has 40 layers / 2560 embedding, the 8B has 40 layers / 4096 embedding, and the 30B has 64 layers / 32768 MLP hidden size. Published context is 128K tokens with a 512K long-context pre-training phase.
  • Pre-training: ~15 trillion tokens from scratch, including 1 trillion tokens of synthetic code generated via IBM's CodeAlchemy pipeline.
  • Post-training pipeline: A multi-stage, multi-environment RL chain using asynchronous GRPO with leave-one-out baselines and truncated importance sampling. The sequence is RLVR → skill boosters → SWE → Terminal → Search → RLHF. SFT used ~7.2M samples (~100B tokens, ~65B trainable) with a 31.6% agentic / 68.4% non-agentic mix, and trajectories were generated across OpenHands, SWE-agent, Terminus-2, MiniSWE, Codex, and Goose harnesses, judged by GPT-OSS-120B and Gemma 4.
  • Agentic RL scope: Only the 8B and 30B models receive the agentic RL block (SWE, Terminal, Search), while the 3B receives foundational RL and alignment only — a key design decision explaining the capability gap across sizes.
  • Granite Speech 5.0 Turbo CTC: 470M-parameter connectionist temporal classification (CTC) model with no LLM backbone, achieving ~12,600 RTFx on a single H200 (vs. ~6,000 for current Open ASR leaderboard leaders). WebGPU demo available.
  • Serving: Speculative decoding layer included for faster inference; GGUF quants down to Q4_K_M released; FP8/NVFP4 support on vLLM for the 30B.

Industry Insight

  • The deliberate exclusion of agentic RL from the 3B model signals that IBM views autonomous tool-use capability as a scale-dependent property, suggesting that smaller reasoning models will remain primarily suited for structured, non-agentic tasks — practitioners should match model size to agent complexity requirements rather than assuming capability scales linearly.
  • The multi-stage GRPO pipeline with leave-one-out baselines and truncated importance sampling offers a replicable blueprint for open-reasoning model development; teams building proprietary reasoning models may adopt similar asynchronous RL chaining to avoid the instability of single-pass RLHF on long-CoT trajectories.
  • The Apache 2.0 licensing combined with on-premise deployment viability positions Granite 4.2 as a strong candidate for regulated industries (financial services, healthcare, public sector) that require both reasoning capability and data sovereignty, potentially accelerating open-weight model adoption in compliance-sensitive verticals.

TL;DR

  • IBM发布Granite 4.2开源推理语言模型家族,包含3B、8B和30B三种参数规模,全面采用Apache 2.0许可证
  • 模型架构为解码器密集Transformer,预训练约15万亿token,支持显式思维链推理及thinking/non-thinking/低开销模式切换
  • 8B和30B版本引入Agentic RL训练阶段,在真实沙盒环境中学习代码编辑、终端操作和网页搜索能力
  • 同步发布Granite Speech 5.0 Turbo CTC语音模型(470M参数),在H200上实现约12,600 RTFx吞吐量,较现有速度领先者提升约一倍
  • 30B模型在SWE-Bench Verified达到57.00、Terminal-Bench 2.1达到29.24,AIME25达到89.17,展现强代码与推理能力

为什么值得看

Granite 4.2将开源推理模型与Agentic RL训练范式结合,为开发者提供可本地部署、可商用的高性能代码智能体基座。其多模式推理开关设计兼顾了推理质量与推理成本,对追求可控AI部署的企业具有直接参考价值。

技术解析

  • 模型架构:解码器-only密集Transformer,采用GQA(8个KV头)、RoPE(θ=10,000,000)、SwiGLU MLP、RMSNorm(ε=1e-5),使用bfloat16精度,支持128K原生上下文(预训练含512K长上下文阶段)
  • 训练流程:预训练15万亿token,SFT使用约720万样本(约100B token,65B可训练),其中31.6%为Agentic数据(软件工程占Agentic数据的69%),质量评估使用GPT-OSS-120B和Gemma 4作为裁判
  • 多阶段RL训练:采用异步GRPO进行多阶段后训练,顺序为RLVR→技能增强→SWE→Terminal→Search→RLHF,使用leave-one-out基线和截断重要性采样控制策略偏移,Agentic RL仅应用于8B和30B
  • 推理优化:提供Speculative Decoding加速推理,支持GGUF量化(低至Q4_K_M),可通过Ollama/LM Studio本地部署,30B支持FP8/NVFP4在vLLM上服务
  • 语音模型:Granite Speech 5.0 Turbo CTC为470M参数CTC模型,无LLM骨干,在Open ASR榜单上RTFx达12,600(H200单卡),约为现有领先者6,000的两倍

行业启示

  • 开源推理模型进入Agentic时代:IBM将Agentic RL引入开源模型训练,标志着开源模型从"被动问答"向"主动执行"演进,企业可基于此构建低成本、可审计的智能体系统
  • 推理成本与质量的权衡设计:thinking/non-thinking/低开销三模式切换为生产环境提供了灵活的推理预算控制策略,适合对延迟和成本敏感的高并发场景
  • 语音-文本融合架构的简化趋势:CTC模型去除LLM骨干直接实现ASR,在吞吐量上实现翻倍,提示语音AI领域可能走向更轻量的专用架构而非通用大模型方案

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Open Source 开源 LLM 大模型 Agent Agent Code Generation 代码生成 Training 训练