Saturated simulation scores mean LIBERO-Pro and Robosuite no longer separate strong methods from weak ones, so real-robot results and continual refinement become the meaningful signals here. Robotics researchers and embodied-AI engineers can clone the code and benchmark their own stacking or manipulation pipelines against it. The GPT-6 Astra integration hints at a general model driving physical manipulation rather than task-specific policies.
面壁智能的 MiniCPM-V 4.7 1B 能直接在 iPhone 上离线读小票答问题，1B 模型本地跑多模态，这个演示挺直观。
微软的语音转写模型 MAI-Transcribe-2 上架 OpenRouter 了，想用 API 做转写的朋友可以直接去试试。
微软的 MAI 转录模型在 OpenRouter 上准确率和调用量都排第一，做语音智能体的可以试试。
Day-0 serving and RL support means practitioners can deploy and fine-tune a 501B-parameter sparse model without waiting for ecosystem catch-up. Teams running large-scale inference can start with vLLM and add reinforcement learning through SkyRL's LoRA or MXFP8 paths. The 1M-token context window puts Beam in direct competition with other frontier sparse models for long-document and agentic workloads.
Gemini 4 Argon 刚拿下 WANDR 纪录，用 Perplexity Pro 或 Max 的话现在就能切换体验，可以去试试搜索效果差多少。
Meta 的 816B 大模型现在有量化版能跑在 256GB 内存的机器上了，224GB 那档保住 89.2% 原始质量，有硬件的可以试试。
Reflection把Beam 1.0开源了，501B总参但只激活23B，跑起来不算重，不过评分比GLM 5.3 Flash低一些，可以拿来对比试试。
Apache 2.0 weights with a 1M context window make Beam one of the few permissively licensed models built explicitly for agentic coding workloads. Developers can run it locally or via Ollama Cloud with zero data retention, and it plugs directly into Claude Code, Codex and OpenCode through Ollama's launcher commands. Teams evaluating open alternatives to closed coding models now have a new option to benchmark against their own agent pipelines.
An Apache 2.0 release with NVFP4 quantised weights and a companion training harness is rare, since most open-weight drops stop at the checkpoint. Teams running inference on NVIDIA hardware can take the DFlash2 drafter and FP8/NVFP4 variants directly into production stacks without negotiating licences. Engineers evaluating open models against Llama or Qwen releases should benchmark Beam on their own workloads.
24 of the top 25 open models on Code Arena: WebDev come from Chinese labs, and Beam 1.0 is the lone US exception — a notable shift in the open-weights competitive landscape. Builders evaluating Apache 2.0 agentic models can compare Beam directly against Qwen3.8, Kimi K2.5 Thinking and GLM-5 once the full weights drop this month. Its 23B active parameter design targets reasoning efficiency for agent workloads.
Usage on OpenRouter is a proxy for real developer demand, and 19.5T tokens in a first week puts Step 5 Preview alongside the most-routed open models worldwide. Teams choosing inference backends can compare its cost and throughput against Qwen and DeepSeek offerings on the same leaderboard. Its traction also signals continued Chinese lab momentum in the global API routing market.
An Apache 2.0 release at 501B scale with day-0 4-bit checkpoints gives enterprises an openly licensed frontier-class option alongside DeepSeek and Qwen. Engineers can download the NVFP4 weights directly from Hugging Face and run them on Blackwell Ultra hardware at the stated throughput. The inclusion of the Mirror CLI training harness also makes the full stack reproducible for teams building their own large models.
Day-one inference support on SGLang means teams can self-host Beam immediately instead of waiting on API-only access, an unusual level of launch coordination between a model vendor and an open inference framework. Practitioners evaluating large models against vLLM or TensorRT-LLM can use this deployment path as a direct comparison point. Watch whether other labs adopt similar launch-day community integrations.
Reflection AI 放出了一个新的开源模型，vLLM 还顺手写了优化教程，想自己部署跑推理的可以看看这篇。
Reflection 的第一个开源模型 Beam 出来了，501B 参数只激活 23B，跑分超过 NVIDIA 和 K2 Horizon，但幻觉率有点高，终端任务也比较弱。
Instead of wrapping a general model in a memory framework like Mem0 or MemOS, this uses a small dedicated writer model, cutting inference cost to roughly 3B-model levels while staying cross-framework. Developers building agent pipelines can pip install it and swap storage backends freely, and Mac users can share memory across coding tools via MemHub. The sub-1-point margin over Claude Opus 5 is self-reported and awaits independent replication.
The sparse-activation design at 501B/23B active with a 1M-token window makes Beam a candidate for long-context workloads without full MoE serving costs. Teams running vLLM can deploy it immediately and fine-tune with SkyRL's LoRA or MXFP8 RL paths. Compared with other large reasoning models, Beam's Day-0 coverage of both serving and RL tooling is the differentiator.
A sparse mixture-of-experts model with an integrated agent harness reduces the token cost of running long coding sessions, which directly lowers API and compute bills. Developers building agentic coding workflows can try it on day one through Transformers and pick NVFP4 or FP8 checkpoints to fit their hardware budget.
Mistral 晒出自家 Large 4 在 Ulam 家 STEMBench 上工程和天文题的成绩，做 STEM 任务的可以看看这个榜单。
The trade-off here is deliberate: Liquid AI swapped raw accuracy for knowing when not to answer, which matters more for agentic decision workflows where a wrong answer is costlier than a refusal. Builders routing high-volume decisions can compare it directly against Celeris-1 Decision ($32 per 1M tasks), Mercury Decide ($15) and GPT-6 Luna (Decisions) ($79) on both price and speed.
腾讯混元开源了个记忆管家 Hy Memory 2.0，30B 模型自称打赢 Opus 5，MemHub 还能在 10 多款工具间同步记忆，搞智能体的可以试试。
有人发帖夸 Claude Opus 5.5 是现在最强的模型，说 Anthropic 找回了状态，想换模型前可以看看这波评价。
Meta 把 816B 的 MoE 模型开源了，还有 NVIDIA 和 FLUX 的新模型，一条看完今天所有大新闻。
快手 Kling 4.0 放出一段一镜到底的游戏概念视频，镜头跟着角色从窄缝飞到开阔天空，可以看看现在视频生成的一镜到底水平。
Replacing a dedicated speech decoder with a general-purpose 4B LLM means recognition inherits stronger word-context modelling, which helps resolve acoustically ambiguous words — an architectural shift other ASR builders may follow. Developers building multilingual voice features can run it locally today under Apache 2.0, inspecting or fine-tuning it without API costs, and benchmark it against Whisper-class models on their own audio.
黑森林开源了 FLUX 3 Image Klein，文生图、改图、多图参考都能做，作者拿它和 Ideogram 4.5、Krea 3 比了效果，还提醒中文字容易翻车。
Xiaomi 的 MiMo V2.6 Pro 在 Agent Arena 开源榜拿了第一，Pro 和 Flash 还都卡在性能成本最优区间，跑智能体的可以关注下。
Muse Spark Open 跑在 vLLM 上了，支持多模态和 1M 上下文，还有个挺可爱的演示小故事，可以看看。
Meta 开源了 816B 的 Muse Spark Open，两台 512GB 的 Mac Studio 加量化就能本地跑，生成速度约 25 Token/s，喜欢本地部署的可以看看这套方案。
Intern-S2 直接参与写数学论文了，北大团队还把 4 条定理全部形式化验证，论文和代码都放出来了，数学方向的人可以看看。
A 30-year-old conjecture falling to a Chinese lab's model plus Lean verification shows AI moving from coding benchmarks into formal mathematics, where every step is machine-checked. Researchers in theorem proving and anyone tracking Chinese labs' frontier ambitions should follow the write-up. It also gives Lean-based formalisation workflows a high-profile proof point against informal LLM math claims.
Same-day vLLM support means teams running RL fine-tuning pipelines can serve and extend Muse Spark Open without waiting for third-party integration. Builders working on open-source model training will want to compare its RL infrastructure setup against GRPO-style workflows on TRL or Ray. The collaboration also signals vLLM consolidating its position as default inference infrastructure for large-scale model training.
Meta 又发开源模型了，816B 参数但只激活 42B，还能吃视频和音频，跑分追平 Qwen3.8，做应用前记得看商用条款。
有消息说智谱的 GLM 5.5 快发布了，靠用户数据做 RL 后训练，写代码有望追上前沿闭源模型，开源党可以关注下。
NVIDIA 新一代 Vera Rubin NVL72 的跑分出来了，跑 DeepSeek V4 Pro 每兆瓦吞吐量是上一代的 5.7 倍，还刷出 906 tok/s 的单用户最快速度。
Only the latest 2000 items are listed; for older ones see the daily archive
