[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-nemotron-bolt-e4b-4gb-edge-slm":3,"topics-all":38,"news-related-265c8b9e-9a65-4198-bc00-a6ba1c8b8248":67},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"265c8b9e-9a65-4198-bc00-a6ba1c8b8248","英伟达开源 4GB 小模型:Bolt E4B 少说话多干活","NVIDIA 开源 Nemotron Bolt E4B:Gemma-4-E4B 后训练的 4GB 边缘小模型,主打语音助手与机器人交互。官方称对话基准平均较上代高约 40%,输出 token 最多省 80%,RTX 5070 上 3.5GB 显存跑 133 token\u002Fs,训练配方全公开。","大模型往小处走，英伟达给了新样本：10 月 23 日，Nemotron 3.5 Bolt E4B 4GB 开源发布。它在 Google Gemma-4-E4B-it 基础上后训练而来，总参 8B、有效 4B，目标不是数据中心，而是 Jetson Orin Nano 8GB、RTX 5070 12GB 这一档——语音助手、游戏 NPC、机器人接口都是点名场景。\n\n## 4GB 里到底装了什么\n\n按官方模型卡，Bolt E4B 主打 Q4_K_M 量化形态，目标负载下显存约 4GB，上下文 64K，支持 35 种语言（底座预训练覆盖 140+）。功能全围绕\"交互智能体\"：多轮工具调用（含异步集成指引）、可开关对话模式（鼓励简短语音式回复）、可配置推理（难题开思考、简单直接答，思考长度可设上限压延迟）。\n\n官方实测（Q4_K_M、批大小 1、llama.cpp）：RTX 5070 12GB 上显存 3.5GB、首 token 548ms、133.4 token\u002Fs；Jetson Orin Nano 8GB 上 4.2GB、15.5 token\u002Fs。官方博客还展示 Reachy Mini demo：Cosmos 3 Edge 供视觉理解，Bolt E4B 管动画脚本和语音回复。\n\n## 少说话，多干活\n\n官方口径的两个数字：对比上代 Nemotron 3 Nano 4B，对话智能体基准平均分高约 40%，比同级对手高约 10%；推理模式下输出 token 比对手最多少 80%。边缘端少生成 token 不只省钱——解码量直接对应延迟和功耗，\"少说话\"本身就是卖点。\n\n压话痨的细节：RLVR+RLHF 双轨（NeMo-RL \u002F NeMo-Gym），每个样本同时以推理\u002F非推理两种模式使用，批次内非推理:推理保持 2:1；训练中答案越拉越长，团队在 GenRM 环境加大组长相对长度惩罚控制冗余。\n\n## 赢在自选战场，输在通用基准\n\n逐行看官方跑分表，方向比总数诚实。赢面集中在对话与人设：MultiChallenge 53.1（Qwen3.5-4B 51.8）、RoleBench 67.0、IHEval 84.8 等项全是四模型最高。但通用 agent 基准不全赢：Tau2-Bench 55.1 明显落后 Qwen 的 75.1，RULER 87.0 排在 Qwen 95.1 之后，EQ-Bench 39.4 略低于原版 Gemma。官方也承认基准按客户负载挑，\"选模型不是挑榜单最高分\"。\n\n## 量化与加速：配方一起放出来\n\n量化路径也公开：BF16 checkpoint 经 llama.cpp 重要性矩阵（imatrix）校准到 Q4_K_M，校准数据约 100 万 token，取自与选中 checkpoint 同一步的 RL rollout。另有 DFlash 并行草稿模型，一次前向提出全部草稿 token，官方称比 Gemma-4-E4B 原生 MTP 吞吐高至多 21%。四组 checkpoint 全放出，Ollama、LM Studio 当天可用，许可证 OpenMDW-1.1。\n\n## 所以呢\n\n边缘小模型的竞争正从\"谁分高\"转向\"每 GB 智能\"。Bolt E4B 的思路：借 Google 底座、用自家后训练配方在 4GB 档做专精，再把配方开源——模型会过时，可复现的配方和量化细节才是留给社区的。输给 Qwen 的基准则提醒：厂商图表永远长在自家主场，你的 workload 才是裁判。\n\n参考：research.nvidia.com\u002Flabs\u002Fnemotron\u002FNVIDIA-Nemotron-3.5-Bolt-E4B-4GB\u002F；huggingface.co\u002Fnvidia\u002FNVIDIA-Nemotron-3.5-Bolt-E4B-4GB","https:\u002F\u002Fresearch.nvidia.com\u002Flabs\u002Fnemotron\u002FNVIDIA-Nemotron-3.5-Bolt-E4B-4GB\u002F","474eef8c-e0c3-46cf-adee-c089558220f9",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"8dac812d-3839-4abe-a855-5f56ec9515fd","nvidia",{"id":19,"name":20,"slug":20,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":22,"name":23,"slug":23,"description":14,"color":14},"045c011e-e2bb-45ce-bdd6-0c927f8a3b87","token-efficiency",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"1f124327-6633-4f3c-809c-c17aaae66f40","en","Bolt E4B 4GB: NVIDIA's 4GB open SLM talks less, does more","NVIDIA open-sources Bolt E4B, a 4GB Gemma-4-E4B-based SLM for edge voice agents: ~40% over Nano 4B, up to 80% fewer tokens, 133 tok\u002Fs at 3.5GB.","NVIDIA's latest answer to \"small models, small memory\" shipped on October 23: Nemotron 3.5 Bolt E4B 4GB, an open small language model post-trained on Google's Gemma-4-E4B-it. It packs 8B total parameters (4B effective) into roughly 4GB of memory in its target workloads, and it is not aimed at data centers — the named targets are Jetson Orin Nano 8GB, RTX 5070 12GB-class edge hardware, voice assistants, in-game NPCs, and robot human-machine interfaces.\n\n## What fits in 4GB\n\nPer the official model card, Bolt E4B ships primarily as a Q4_K_M quantized release: ~4GB memory in target workloads, 64K context, 35 languages supported (the base was pre-trained on 140+). Every feature is built around interactive agents: multi-turn tool use with asynchronous-call integration guidance, a toggleable conversational mode that encourages concise speech-ready replies, and configurable reasoning — thinking on for hard requests, off for fast ones, with a cap option to meet latency targets.\n\nOfficial measured numbers (Q4_K_M, batch size 1, llama.cpp) tell the positioning story: 3.5GB VRAM, 548ms time to first token and 133.4 tokens\u002Fs on an RTX 5070 12GB; 4.2GB and 15.5 tokens\u002Fs on a Jetson Orin Nano 8GB. The blog also demos a Reachy Mini robot on Jetson Orin Nano 8GB, where Cosmos 3 Edge supplies real-time visual understanding while Bolt E4B drives animation scripts and spoken responses.\n\n## Talks less, does more\n\nTwo vendor numbers stand out: roughly 40% higher average quality than the prior Nemotron 3 Nano 4B on conversational-agent benchmarks (about 10% over other peers evaluated), and up to ~80% fewer output tokens than peers in reasoning-enabled comparisons. On edge devices, generating fewer tokens is not just cost math — decoding work maps directly to latency and power, which is why brevity is a feature here.\n\nThe post-training recipe behind that brevity is documented: RLVR plus RLHF via NeMo-RL and NeMo-Gym, with every training sample used in both reasoning and non-reasoning modes at a 2:1 non-reasoning-to-reasoning ratio per batch. As answers grew longer during training, the team increased group-relative length penalties in the GenRM environment to control verbosity.\n\n## Wins on its home field, loses on general benchmarks\n\nRead the official benchmark tables row by row and the direction is more honest than the headline. The wins cluster in conversation and persona: MultiChallenge 53.1 (vs Qwen3.5-4B 51.8, Gemma-4-E4B 42.1), RoleBench 67.0, RoleMRC 80.6, instruction-hierarchy IHEval 84.8, When2Call-MT 63.4 and LongMemEval 72.1 — all best among the four compared models. But it does not sweep general agent benchmarks: Tau2-Bench 55.1 clearly trails Qwen3.5-4B's 75.1, BFCL-v3 68.5 sits below Qwen's 70.9, RULER 64K at 87.0 ranks behind Qwen's 95.1 and even Nano's 91.5, and EQ-Bench 39.4 is slightly below the original Gemma-4-E4B's 41.1. NVIDIA itself concedes the benchmarks were chosen to reflect customer workloads — \"choosing a model is not as simple as picking the highest leaderboard score.\" Read the trade-offs before deploying.\n\n## Quantization and speed: the recipe ships too\n\nThe BF16 checkpoint was quantized to Q4_K_M with llama.cpp importance-matrix (imatrix) calibration, using roughly 1M tokens of RL rollouts from the same training step as the selected checkpoint — calibration data that tracks the training run is a reusable engineering detail. The release also includes a DFlash parallel drafter that proposes all draft tokens in one forward pass; NVIDIA reports up to 21% higher throughput versus Gemma-4-E4B's original MTP drafter. Four checkpoints (GGUF, DFlash, BF16, BF16-DFlash-NVFP4) are out, available day-one on Ollama, LM Studio, Unsloth and the Jetson AI Lab, under the OpenMDW-1.1 license.\n\n## So what\n\nEdge-model competition is shifting from \"who scores highest\" to intelligence per GB and tokens per second. Bolt E4B's play: take Google's base, specialize with NVIDIA's post-training recipe at the 4GB tier, then open-source the recipe itself — models age, reproducible recipes and quantization details are what stay. And the benchmarks it loses to Qwen are a useful reminder: vendor charts always grow on their home field; your workload is the only referee.\n\nRefs: research.nvidia.com\u002Flabs\u002Fnemotron\u002FNVIDIA-Nemotron-3.5-Bolt-E4B-4GB\u002F ; huggingface.co\u002Fnvidia\u002FNVIDIA-Nemotron-3.5-Bolt-E4B-4GB","nemotron-bolt-e4b-4gb-edge-slm","2026-10-24T21:09:01Z","2026-10-24T21:09:04.354454Z","2026-10-24T21:09:04.354462Z",true,"agent",606,[39,49,58],{"slug":40,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":35,"created_at":47,"modified_at":48},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":50,"tag_slug":50,"title_zh":51,"title_en":52,"intro_zh":53,"intro_en":54,"id":55,"is_active":35,"created_at":56,"modified_at":57},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":59,"tag_slug":59,"title_zh":60,"title_en":61,"intro_zh":62,"intro_en":63,"id":64,"is_active":35,"created_at":65,"modified_at":66},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":68},[69,74,79,84,89,94],{"id":70,"title":71,"news_slug":72,"published_at":73},"039d2494-af49-4d5d-a89f-e3ca91f60df5","转写没有标准答案:英伟达开源43亿激活语音大模型","nvidia-nemotron-3-5-transcribe","2026-10-29T15:15:27+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"2ddb354a-06f7-4781-b2d5-bcfd1faf38a5","Dory 1.0:英伟达4B循环模型,指令遵循反超9B","nvidia-dory-1-0-4b-looped-llm","2026-10-20T21:09:38+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"d72b4ca8-008c-4bc7-b527-e9d131d6a379","IBM Granite Speech 5.0:把 ASR 里的语言模型砍掉,470M 跑出 3.5 小时\u002F秒","ibm-granite-speech-5-turboctc-470m","2026-09-02T05:06:25+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"de2cceb2-7d39-4a5f-844e-5a3144667f49","Nemotron 3.5 Lightning 开源：30B 总参 3B 激活的混合 MoE，直接用 NVFP4 配方预训练","nemotron-35-lightning-30b-a3b-open-release","2026-08-16T15:00:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"41c8976c-d775-4603-aa01-693c57b7b0bd","NVIDIA Nemotron 3 Embed 登顶 RTEB：把 8B 旗舰检索能力蒸馏进 1B 部署款","nvidia-nemotron-3-embed","2026-07-16T18:00:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"b446777f-9949-47e4-8ca8-2d0fa7f14126","NVIDIA Cosmos 3 Edge 4B开源世界模型：物理AI实时推理从云端搬到产线","nvidia-cosmos-3-edge-4b","2026-07-16T14:00:00+00:00"]