[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-servicenow-apriel-voice-27b-preview":3,"topics-all":38,"news-related-34bd9f83-04c2-4ffa-8e68-f65a74dfb490":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},"34bd9f83-04c2-4ffa-8e68-f65a74dfb490","开源语音智能体破50分:Apriel Voice 27B","ServiceNow开源语音模型Apriel Voice，用Whisper编码器对接Gemma 4底座，直接听音频、吐出文字和工具调用。官方评测里，它的服务准确率51.27分，对话体验74.13分为全场最高，任务成功率仍落后闭源18.68分。权重按MIT许可放出，四张H100就能跑起来。","开源语音智能体一直卡在个位数——ServiceNow 自己的评测表里，UltraVox v0.6 的服务准确率只有 8.94 分，Audex-30B-A3B 只有 5.66 分。10 月 29 日挂上 Hugging Face 的 Apriel Voice 27B Preview（[论文](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.40738)），把开源阵营一把拉到 51.27 分。它的定位一句话讲完：直接听音频，吐出文字回复和结构化工具调用。\n\n## 两段式打法：听得懂，但不负责说\n\n现在的语音助手基本两条路：要么 ASR+LLM+TTS 三段拼接，要么端到端的语音到语音（S2S）。Apriel Voice 选了中间的两段式：一个 Whisper large-v3-turbo 语音编码器，通过学习出来的投影层，接在 Gemma 4 26B-A4B 后训练底座上。模型吃 16kHz 音频，输出文本和工具调用；语音合成交给外部——官方评测里所有两段、三段系统统一用 Cartesia Sonic 3.5 合成，闭源 S2S 用自家合成。模型卡写得直白：这个 checkpoint 本身不做全双工语音。\n\n两段式的赌注很清楚：语音接口只管「听」，推理和工具调用全部复用现成的开源底座。这样语音能力可以跟着底座升级走，不用每次为语音重训整个模型；代价是「说」这件事永远依赖外部合成器，延迟和音色都不归自己管。\n\n训练分两阶段：第一阶段建语音接口，第二阶段投影层与底座联合适应，用监督微调、文本回放（text replay）、按数据源切换的交叉熵与前向 KL 监督；对话数据靠自动构造的语音条件对话生成，蒸馏走在线路线——本地教师服务、师生计算重叠、融合全词表损失。\n\n## 成绩单：体验分领先，成功率还差一截\n\nthinking 模式下：Big Bench Audio 98.90、Speech-BFCL 81.13、τ-Voice 任务成功率 65.51%，平均延迟 1.91 秒；八项音频推理基准均值 85.84。转写侧，LibriSpeech 词错率 test-clean 2.23%、test-other 5.15%，GigaSpeech 10.20、μ-Bench（en-US）10.76、FLEURS（en-US）3.55、Earnings21 11.25，越低越好。\n\n对标闭源 S2S：GPT Realtime 2.1 服务准确率 59.03、对话体验 61.6、τ-Voice 69.82、延迟 1.55 秒；Gemini-3.8 Live Extended Thinking 是 54.82、73.5、84.19、1.50 秒。在这张官方评测表里，Apriel 的对话体验 74.13 是全场最高——比 Gemini 高 0.63 分，比 GPT Realtime 高 12.53 分；但任务成功率 65.51 落后 Gemini 18.68 分，服务准确率也还差 GPT Realtime 7.76 分，延迟慢 0.41 秒。服务准确率量的是订没订对，对话体验量的是聊得顺不顺——两个指标的裂口说明：它强在对话节奏，执行还没跟上。一句话：聊得像回事了，活儿还没干利索。\n\n## 加了语音，文字能力基本没丢\n\n对比训练前后：MMLU-Pro 85.27→84.42 微降，GPQA Diamond 77.80→80.40 反升，AIME 2024\u002F2025 持平在 92.00\u002F81.67，MATH-500 92.76→92.88，智能体基准 τ²（三个域均值，拿 GPT-5.2 当用户模拟器）68.93→70.95。文本回放这次显然起了作用——语音接口没有吃掉推理底座。\n\n## 部署门槛与许可\n\n权重 27B、BF16，MIT 许可。官方提供 vLLM v0.32.0 Docker 镜像，参考配置四张 H100 80GB（tensor-parallel 4）、262144 上下文，投机解码拿 gemma-4-26B-A4B-it-assistant 当草稿模型；优化镜像在四张 H100 上实测输出吞吐约为基线配置的 2 倍。长音频按 30 秒分块，官方称 30 分钟内表现良好，一小时输入要把准入上限调到 3600 秒、批预算 32768。仓库近一个月下载 338 次，作者本人今天才把论文提交到社区榜，目前 16 个 upvote——这个模型还很新。\n\n## 官方自己泼的冷水\n\n模型卡的 Limitations 写得不含糊：多语言表现不均，英文成绩不代表其他语言；噪声与远场语音仍是难点；复杂工具工作流的忠实度和可靠完成仍是限制；长录音可能复读或失去连贯；跑分只描述被评测的配置，换个 workload 不保证成立。另外，全部数字来自 ServiceNow 单一发布方的自报评测，暂无第三方复现；「Preview」这个后缀名副其实。\n\n开源阵营在这张评测表上站上了 50 分档，「编码器+开源推理底座+投影层」的两段式路线被验证成立。但 84.19 分的任务成功率和 1.5 秒延迟还是闭源 S2S 的护城河。听得懂之后，下一个战场是做得成——你的语音助手，聊得再好，订不到那趟航班就还是零分。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.40738","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"c4455801-0273-477d-9e92-00b4addfa439","asr",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"fbe94af7-af0e-4821-b411-bd312352a3f2","en","Apriel Voice 27B: Open Voice Agents Cross the 50-Point Line","Apriel Voice 27B takes speech in and returns text plus tool calls; on ServiceNow's own EVA eval its 74.13 experience score tops all systems.","Open-source voice agents have been stuck in the single digits — in ServiceNow's own evaluation table, UltraVox v0.6 scores just 8.94 on service accuracy and Audex-30B-A3B only 5.66. On Oct 29, ServiceNow put Apriel Voice 27B Preview ([paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.40738)) on Hugging Face and pulled the open camp straight to 51.27. The pitch in one sentence: it takes audio in and returns text replies plus structured tool calls.\n\n## Two stages: it listens, but does not speak\n\nToday's voice assistants mostly follow one of two routes: stitch ASR + LLM + TTS together, or go end-to-end speech-to-speech (S2S). Apriel Voice takes the middle path — two stages. A Whisper large-v3-turbo speech encoder connects, through a learned projector, to the post-trained Gemma 4 26B-A4B backbone. The model ingests 16kHz audio and outputs text and tool calls; speech synthesis is left to an external component. In the official evaluation all two- and three-stage systems use Cartesia Sonic 3.5, while proprietary S2S systems use their own. The model card says it plainly: the checkpoint itself does not do full-duplex speech.\n\nThe bet behind two stages is explicit: the speech interface only listens, while reasoning and tool calling reuse an existing open backbone. Speech capability can then ride backbone upgrades instead of retraining a whole model per modality; the price is that speaking forever depends on an external synthesizer, with latency and voice out of the model's hands.\n\nTraining runs in two stages. Stage one establishes the speech interface; stage two jointly adapts the projector and backbone with supervised fine-tuning, text replay, and source-dependent cross-entropy plus forward KL supervision. Conversational data comes from automatically constructed speech-conditioned dialogues, and distillation is online — local teacher serving, overlapping teacher and student computation, and fused full-vocabulary losses.\n\n## The score sheet: experience leads, task success lags\n\nIn thinking mode: 98.90 on Big Bench Audio, 81.13 on Speech-BFCL, 65.51% task success on τ-Voice with 1.91 seconds mean latency, and a mean of 85.84 across eight audio-reasoning benchmarks. On transcription, LibriSpeech word error rate is 2.23% on test-clean and 5.15% on test-other, with GigaSpeech at 10.20, μ-Bench (en-US) at 10.76, FLEURS (en-US) at 3.55 and Earnings21 at 11.25 — lower is better.\n\nAgainst proprietary S2S: GPT Realtime 2.1 sits at 59.03 service accuracy, 61.6 experience, 69.82 τ-Voice, 1.55 seconds latency; Gemini-3.8 Live Extended Thinking at 54.82, 73.5, 84.19, 1.50 seconds. In this official table, Apriel's 74.13 experience score is the highest of any system — 0.63 above Gemini and 12.53 above GPT Realtime. But its 65.51 task success trails Gemini by 18.68 points, service accuracy still sits 7.76 behind GPT Realtime, and latency is 0.41 seconds slower. Service accuracy measures whether the booking is right; experience measures whether the conversation flows. The gap between the two says it plainly: strong on conversational rhythm, execution still lagging. It converses well; the job is not yet finished.\n\n## Speech came in; text ability mostly stayed\n\nComparing before and after speech training: MMLU-Pro 85.27→84.42 (slightly down), GPQA Diamond 77.80→80.40 (up), AIME 2024\u002F2025 unchanged at 92.00\u002F81.67, MATH-500 92.76→92.88, and the τ² agent benchmark (mean across three domains, with GPT-5.2 as the user simulator) 68.93→70.95. Text replay clearly did its job — the speech interface did not eat the reasoning backbone.\n\n## Deployment and license\n\nThe checkpoint is 27B params in BF16 under an MIT license. ServiceNow ships a vLLM v0.32.0 Docker image; the reference setup runs on four H100 80GB GPUs (tensor-parallel 4) with a 262144 context window, and speculative decoding uses gemma-4-26B-A4B-it-assistant as the draft model. On four H100s the optimized image delivers roughly 2x the measured output throughput of the baseline configuration. Long audio is split into 30-second chunks, with good performance reported up to about 30 minutes; one-hour input requires raising the admission limit to 3600 seconds and the batch budget to 32768. The repo shows 338 downloads in the past month, and the author himself submitted the paper to the community board only today — it currently has 16 upvotes. This model is very new.\n\n## The cold water, from the vendor itself\n\nThe model card's Limitations section does not hold back: multilingual performance is uneven, and English results do not imply other languages; noisy and distant-microphone speech remain challenging; faithfulness and reliable completion of complex tool workflows are still limited; longer recordings can repeat themselves or lose coherence; and the benchmark scores describe the evaluated configurations only. Every number also comes from a single self-reported evaluation by ServiceNow, with no third-party replication yet — the \"Preview\" suffix means what it says.\n\nThe open camp now stands in the 50-point band on this evaluation table, validating the two-stage route of encoder plus open reasoning backbone plus projector. But an 84.19 task success rate and 1.5-second latency remain the proprietary moat. Once listening is solved, the next battlefield is getting the job done — a voice assistant that converses beautifully but fails to book the flight still scores zero.","servicenow-apriel-voice-27b-preview","2026-11-02T13:19:00Z","2026-11-02T13:20:07.002074Z","2026-11-02T13:20:07.002086Z",true,"agent",394,[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},"b6d8130b-f989-46a5-baa3-d81e66265ef8","Reflection开源Beam:501B MoE的RL账本","reflection-beam-501b-moe-rl-campaign","2026-11-20T13:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"a79651db-8451-4ee7-9091-809dff098c19","英语只占41%:Translated开源多语MoE小模型","lara-mini-base-multilingual-moe","2026-11-14T19:30:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"503fc4f0-755d-43f1-89cc-8d86b4cdaed9","7.6M参数土耳其语TTS:WER 1.09%却没人做过盲听","antalia-2-mini-turkish-tts-synthetic-data","2026-11-12T01:40:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"ee9161d1-432e-47be-9c9d-84cc76ae3cb9","决策模型当基座:Drex四兄弟全开源,183毫秒出答案","drex-open-source-decision-models","2026-11-06T23:11:15+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"554f6552-6c09-401a-99e3-c9ce60a497ca","ERNIE 5.5 基座:280B激活14B,对话版还没来","baidu-ernie-5-5-flash-base-open-weights","2026-11-05T12:00:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"4193fdcb-58a9-484e-9bd7-e4071c6588bd","Ripple-65B:65B稠密模型80.52分压过975B","ricursive-ripple-65b-dense-open-weight","2026-11-05T05:30:00+00:00"]