2026-10-29开放权重Open weights
Mistral Large 4:1.05T / 49B 激活Mistral Large 4: 1.05T total / 49B active
欧洲开放权重旗舰。1T 级原生多模态 MoE,1M 上下文,128 路由专家(每 token 激活 4 个 + 1 共享),160+ 语言;官方称 3,800 张 Grace Blackwell 训练。A European open-weight flagship: 1T-class natively multimodal MoE, 1M context, 128 routed experts (4 active per token plus 1 shared), 160+ languages; trained on 3,800 Grace Blackwell GPUs per the model card.
- 总参 / 激活Total / active
- 1.05T / 49B(含嵌入输出 52B)
- 专家 / expertsExperts
- 128 routed + 1 shared;4 active
- 上下文 / contextContext
- 1M tokens
- 输入 / inputInput
- 文本 + 图像
- 许可Licence
- Mistral Ecosystem Open-Weight License(修改版 MIT;MaaS 需单独授权)
- 公开指数 / indexPublic index
- AA Coding Agent 48%;AA Intelligence v5 26(第三方,非本站实测)
为什么值得看Why it matters它把「1T 级 MoE」与「1M 上下文」放进开放权重,同时保留欧洲许可叙事与多语言覆盖。对需要在本地或私有云部署旗舰模型、又不想完全依赖闭源 API 的团队,这是 2026 年最重要的新增选项之一。It puts a 1T-class MoE and 1M context into open weights while keeping a European licensing story and multilingual coverage. For teams that need a flagship on-prem or in a private cloud without a fully closed API, this is one of the most important 2026 additions.
边界 / 它没主张什么Boundary / what it does not claim49B 是每 token 激活,不是总计算量;1T 总参意味着权重加载本身就很重,官方给了 NVFP4 档与 DSpark 草稿头,但本地运行仍需多卡。许可是修改版 MIT,不是 OSI 标准;MaaS 商用要单独谈。AA 分数是第三方指数,不能替代你自己的任务评测。49B is active per token, not total compute; 1T total means loading the weights is itself heavy. The release includes an NVFP4 checkpoint and a DSpark draft head, but local hosting still needs multiple GPUs. The licence is modified MIT, not OSI-approved; MaaS use needs a separate grant. AA scores are third-party indices and cannot replace your own task evaluation.
来源Sources:Hugging Face 模型卡Hugging Face model card
2026-10-15开放权重Open weights
Meta Muse Spark Open:816B / 42B 激活Meta Muse Spark Open: 816B / 42B active
Meta 在 Muse 线上的新旗舰,面向多文件重构与长程调试的编程 Agent;816B 总参、42B 激活、1M 上下文、原生多模态,Muse Model License。A new flagship on Meta’s Muse line, aimed at coding agents doing multi-file refactoring and long debugging; 816B total, 42B active, 1M context, natively multimodal, under the Muse Model License.
- 总参 / 激活Total / active
- 816B / 42B
- 上下文 / contextContext
- 1M
- 模态 / modalityModality
- 原生多模态
- 许可Licence
- Muse Model License(厂商许可,非 OSI)
- 本地演示 / local demoLocal demo
- 可在两台 M5 Ultra Mac Studio 运行(AMD 技术文章口径)
为什么值得看Why it matters它把「开放权重」推到 800B 级,并明确以编程 Agent 为主场景;对需要长上下文代码库理解、又希望保留权重控制权的团队,是一个新的上限选项。It pushes open weights into the 800B class and explicitly targets coding agents; for teams that need long-context repository understanding while retaining weight control, it is a new upper bound.
边界 / 它没主张什么Boundary / what it does not claim站点尚未收录它的量化系列,家族卡只记代次与规格;「两台 M5 Ultra 可跑」是厂商演示,不等于通用生产吞吐。Muse Model License 有使用范围限制,商用前读原文。816B 总参的显存门槛极高。This site does not yet cover its quantisations; the family card records only the generation and specs. “Runs on two M5 Ultra Mac Studios” is a vendor demo, not general production throughput. The Muse Model License has scope limits; read it before commercial use. 816B total is an extremely high memory barrier.
来源Sources:AMD 技术文章AMD technical article
2026-09-22开源Open source
小米 MiMo-V2.6:Pro / Flash 双线Xiaomi MiMo-V2.6: Pro / Flash line
Pro / Flash 两条线同时发布并开源;Pro 在 AA Intelligence Index 46.32,1M 上下文,原生全模态(文本 / 图像 / 视频 / 音频),权重与技术报告开放。Pro and Flash launched together with open weights; Pro scores 46.32 on the AA Intelligence Index, with 1M context and native omni-modality (text / image / video / audio); weights and a technical report are open.
- 版本 / variantsVariants
- MiMo-V2.6-Pro / Flash
- 上下文 / contextContext
- 1M
- 模态 / modalityModality
- 文本 / 图像 / 视频 / 音频
- 公开指数 / indexPublic index
- AA Intelligence 46.32(第三方)
- 许可Licence
- 以官方模型卡为准
为什么值得看Why it matters它是 2026 年少数把「原生全模态」与「1M 上下文」同时放进开源权重的中文模型线;对多模态 Agent、视频理解与中文场景,值得单独做一次任务评测。One of the few 2026 Chinese open-weight lines to combine native omni-modality with 1M context; worth a task-specific evaluation for multimodal agents, video understanding and Chinese workloads.
边界 / 它没主张什么Boundary / what it does not claimAA 分数是第三方指数,不覆盖你的领域;「全模态」指输入模态,不代表每一种模态都达到同等质量。Flash 与 Pro 的能力/成本差要分开测。AA is a third-party index and does not cover your domain; “omni-modal” refers to input modalities, not equal quality on each. Test Flash and Pro separately for capability and cost.
来源Sources:小米 MiMo 官方页Xiaomi MiMo official page · Hugging Face collectionHugging Face collection
2026-09-10开放权重Open weights
DeepSeek V4.1 Flash:552B MoE,KV 缓存大幅压缩DeepSeek V4.1 Flash: 552B MoE, much smaller KV cache
552B MoE,Causal-Encoder-Decoder 架构;输入侧 8B 激活、输出侧 16B 激活;原生多模态;官方称 KV 缓存 HBM 占用降至 1/4、SSD 降至 1/8;权重 MIT。552B MoE with a causal encoder–decoder architecture; 8B active on the input side and 16B on the output side; native multimodality; the official post says KV-cache HBM use drops to one quarter and SSD to one eighth; MIT weights.
- 总参 / totalTotal
- 552B MoE
- 激活 / activeActive
- 输入 8B / 输出 16B
- 上下文 / contextContext
- 1M(官方 API 口径)
- KV 缓存 / KV cacheKV cache
- HBM 1/4、SSD 1/8(官方口径)
- 许可Licence
- MIT
- APIAPI
- deepseek-flash;1017 版本 2026-10-17
为什么值得看Why it matters它把「长上下文」的成本从「堆显存」转向「压缓存」,并用 MIT 许可开放权重;对做长文档、代码库与多轮 Agent 的团队,V4.1 Flash 是一个成本敏感型基线。It shifts long-context cost from “more HBM” to “smaller cache” and releases weights under MIT; for long documents, repositories and multi-turn agents, V4.1 Flash is a cost-sensitive baseline.
边界 / 它没主张什么Boundary / what it does not claim1/4、1/8 是官方对缓存占用的口径,不是端到端吞吐提升;实际速度取决于框架、batch 与 SSD。原生多模态的能力边界要按任务测,不能从参数表推断。1017 是小版本更新,接口与行为可能与此前版本有差异。The 1/4 and 1/8 figures are the official cache-footprint claims, not end-to-end throughput gains; real speed depends on the stack, batch and SSD. Native multimodality must be tested per task, not inferred from a spec table. 1017 is a point release; interfaces and behaviour may differ from earlier versions.
来源Sources:DeepSeek 官方发布(2026-09-10)DeepSeek release (2026-09-10) · DeepSeek 1017 更新DeepSeek 1017 update · Hugging Face 权重Hugging Face weights
2026-08-14开放权重Open weights
GLM-5.3:同基座后训练 ScalingGLM-5.3: post-training scaling on the same base
基于 GLM-5.2 同一基座做后训练 Scaling,1M 上下文,官方定位编程与网络安全;权重在发布两周后开放。GLM-5.3-Flash 320B-A18B 于 2026-08-26 上线并以 MIT 开源。Post-training scaling on the same GLM-5.2 base, 1M context, officially positioned for coding and cybersecurity; weights opened two weeks later. GLM-5.3-Flash 320B-A18B launched on 2026-08-26 under MIT.
- 基座 / baseBase
- GLM-5.2 同基座
- 上下文 / contextContext
- 1M
- 定位 / positioningPositioning
- 编程、网络安全(官方口径)
- Flash / 开源Flash / open
- 320B-A18B;MIT;2026-08-26
为什么值得看Why it matters它说明「不换基座、只做后训练」也能把能力推上去,这对算力有限、但已有基座的团队很有参考价值;Flash 档把 320B-A18B 拉到可部署区间。It shows that post-training alone, without a new base, can move capability — useful for teams with a fixed base and limited compute; the Flash tier brings 320B-A18B into a deployable range.
边界 / 它没主张什么Boundary / what it does not claim「编程 / 网络安全」是官方定位,不是第三方基准结论;后训练 Scaling 的收益依赖数据与奖励,不可无条件复制。Flash 是另一套规模,不能把 Pro 的分数直接套到 Flash。“Coding / cybersecurity” is official positioning, not a third-party benchmark result; post-training gains depend on data and rewards and cannot be copied unconditionally. Flash is a different scale; do not transfer Pro scores to it.
来源Sources:智谱研究页Zhipu research page · BigModel 更新日志BigModel release notes
2026-07-27开放权重 / 校正Open weights / correction
Kimi K3:开放日校正为 2026-07-27Kimi K3: open-weights date corrected to 2026-07-27
本站旧卡把 Kimi K3 的开放日写成 2026-06,官方页为 2026-07-27;2026-11 更新把家族卡与图表一起校正。规格不变:2.8T 总参 / 104B 激活,16/896 专家,1M 上下文,原生视觉。The older card dated Kimi K3 to June 2026; the official page says 2026-07-27. This round corrects the family card and the charts. Specs unchanged: 2.8T total / 104B active, 16 of 896 experts, 1M context, native vision.
- 总参 / 激活Total / active
- 2.8T / 104B
- 专家 / expertsExperts
- 896,激活 16
- 上下文 / contextContext
- 1M
- 许可Licence
- Kimi K3 License
为什么值得看Why it matters日期错一个月,会让「谁先发布」的判断整体错位;这类时间线错误比参数错误更隐蔽。校正后,2026 年 7 月的开放权重节点才完整。A one-month date error shifts the “who shipped first” judgement; timeline errors are more hidden than parameter errors. With the correction, the July 2026 open-weight record is complete.
边界 / 它没主张什么Boundary / what it does not claimKimi K3 License 不是 OSI 标准许可;「world’s first open 3T-class model」是厂商自评口径。日期校正只改时间,不改能力结论。The Kimi K3 License is not OSI-approved; “world’s first open 3T-class model” is vendor wording. The correction changes the date, not the capability conclusion.
来源Sources:Kimi 官方发布Kimi official release · 技术报告Technical report