Using egocentric human video instead of robot demonstrations is a data-scaling bet that could cut training costs across humanoid and manipulation startups. The open question is how well skills transfer from human hands to robot hands at deployment. Robotics researchers and hardware teams comparing manipulation models should track whether this approach holds up on real hardware.
The 35B-A3B MoE-style variant with only ~3B active parameters targets on-device multimodal inference, making it a candidate for deployment where cloud APIs are costly or unavailable. Developers in Southeast Asia building offline or low-bandwidth vision-language applications can download weights directly from Hugging Face under open-source terms and benchmark it against Qwen-VL alternatives.
OpenBMB 的 MiniCPM-V 4.7 现在能在 Mac 上离线跑多模态了，1B 版每秒 315 token，图像视频都能看，数据不出本地。
MiniCPM-V-4.7-35B-A3B 出了 20GB 量化版，性能只掉 5% 左右，消费级显卡就能本地跑视觉模型了。
The key change is deployment economics: a 35B MoE now fits consumer devices at usable speeds, because LES hides IO wait through scheduling instead of trading off precision. Developers building privacy-sensitive on-device apps should test it, since no cloud call is required. Against typical acceleration methods that cut quantisation levels and lose points, this keeps baseline accuracy intact.
有推特网友猜测 DeepSeek 的新模型用 RTX 5090 就能跑，如果属实对本地部署党很友好，先看看信息。
State-space models have lagged behind Transformers in inference infrastructure, and this closes part of that gap by cutting memory traffic during Mamba2 decoding. Engineers serving hybrid or SSM-based models in vLLM can enable Mamba2 decode and FlashInfer MTP today. Tri Dao's involvement signals continued convergence of research kernels and production serving stacks.
OpenBMB 的 MiniCPM-V-4.7（35B 总参、3B 激活）跑进了真实 agent 工作流，重点考验工具调用和长上下文，关注高效模型跑 agent 的可以看看。
Joma 拍了支恶搞 GPT-6 Astra 广告的片子，连剪辑署名都写的是模型自己，喜剧人画风，轻松下饭。
有人给不存在的 GPT-6 拍了个洗碗广告恶搞视频，拿"洗碗基准测试"调侃各家跑分的营销味儿，两分钟乐一下。
petergostev 把 Meta 的 Muse Wonder 拉来和 Opus 5.5、GPT-6.1 Sol 这几个模型同台比 3D 场景和前端生成，还附了每次生成的费用，想选模型可以看看。
Meta 新出的 Muse Wonder 直接叫板 GPT-6 和 Claude Opus 5.5，超长程编程跑分高出一截，还能自己扛住长任务不中断，做 Agent 的可以看看跑分细节。
Meta 新发的 Muse Wonder 不按聊天模型做，专攻能替你跑完长任务的 Agent，还自建了两个基准来测'不半途而废'，思路挺不一样。
Muse Wonder 实测不如 Fable 和 Astra，官方说新基准测试很快放出，做视频生成的可以蹲一下细节。
Scale AI 发了个叫 MUSE WONDER 的模型，专门给个人智能体用，卖点是真的能记住你几周几个月的事，金融法律领域还是 SOTA。
Meta AI 负责人 Alexandr Wang 晒出对某个模型的实测感受，说它几乎没短板，其他顶级模型还做不了它能做的事。
Meta 新发了 Muse Wonder 系列模型，主打个人和专业场景的 Agent 任务，Zuckerberg 亲自官宣，做 Agent 的可以关注。
Meta 的 Muse Wonder Full 在 AA Intelligence Index 拿了 56 分世界第二，只差 Claude Opus 5.5 两分，就看什么时候能放出来用了。
Meta 新模型 Wonder Full 离 Opus 5.5 只差 2 分，长代码项目还反超了，做开发的可以去 Muse Code 里试试。
有个叫 SubQ Snap 的决策模型，专门干 agent 工作流里的判断活，384K 上下文，输入才 $0.05/MM，比用前沿模型便宜太多了。
OpenBMB 的 MiniCPM-V 4.7 35B-A3B 被第三方去掉了审查，拒绝率从 96.3% 降到 0，Hugging Face 上能直接下载本地跑。
Meta 放出了 Muse Spark 1.3 的跑分，重点是官方说没刷榜，实际用起来怎么样可以自己试试。
有人提前玩了 Meta 的新模型 Watermelon，说它写代码能自己跑自己查，测试基准居然还超出预期，就是文笔有点干。
OpenBMB 的 MiniCPM-V 4.7 在 3D 生成上赢了 Qwen 3.6 35B-3A，还省 10% token，本地跑视觉模型可以关注一下。
Muse Wonder 出现在 arena.ai 了，可以匿名跟其他模型对战投票，还能看它的 Agent 模式干活。
Meta 弄出了 Fable Class 级模型，两强竞争变三强了，作者还预测 OpenAI 会跟进，看看他怎么说。
Meta 的新模型 Watermelon 要来了，据说是 Fable 级的前沿模型，一边做消费端的 Muse 一边冲前沿，还能不能撑住两线作战挺值得看的。
The benchmark penalises behaviours conventional evals ignore, such as oversized diffs, out-of-scope file edits and unhelpful tests, so its ranking offers a different picture of coding ability than SWE-style test suites. Even the best model fails over 40% of tasks, meaning maintainer cleanup is still unavoidable. Teams choosing models by cost-performance should note GPT-6.1 Sol matching Claude Fable 5.1 at under a tenth of the per-task price.
The benchmark's merge-focused grading penalises the messy diffs and unhelpful tests that cheaper evaluations ignore, so leaderboard gaps here reflect real maintainer acceptance rather than raw test-passing. Leading models still fail over 40% of tasks, and nine tasks went unsolved by every model evaluated. Teams choosing coding agents can weigh GPT-6.1 Sol's 49% score at US$0.81 per task against Claude Opus 5.5's 52% at far higher cost, with open-weights options like MiMo-V2.6-Pro at 45% close behind.
Scale AI 的 Alexandr Wang 实测了 Meta 新模型，说它跟 astra、fable 和 Opus 打平甚至更强，还断言 Meta 会赢下消费级智能体市场。
The finding suggests base model quality, not more post-training, drives generalisation to tasks outside the training distribution — a useful signal for teams deciding where to spend compute. Researchers evaluating reasoning models can use this framing to compare candidates on out-of-distribution scaling curves rather than benchmark scores alone. No regional angle exists; the most concrete comparison is Muse Wonder Full against the Muse Spark 1.3 trend line.
Alexandr Wang 亲口辟谣：muse wonder 会做 3D demo 不是因为专门训练过，他还鼓励大家拿它写真实代码试试。
LMArena 给出实测数据：Kimi K3 和 DeepSeek V4.1 Flash 用不到一成的成本追平了 Claude Fable 5，做 agent 选型前可以看看这份榜单。
Meta 放了两个新模型，主打干活的智能体任务，现在就能在 Agent Arena 里亲手测，排行榜分数还没出。
Meta 新模型 Muse Wonder 发布了，这次用自家的 PSI 指数来评估个人智能体表现，连'大模型味'都量化了。
Scale AI 的 Alexandr Wang 亲自体验了 Muse Wonder 的视频生成，说光照和细节已经能跟 Opus 5.5 掰手腕，可以看看他贴的样片效果。
Only the latest 2000 items are listed; for older ones see the daily archive
