[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-ling-3-tiny-kda-mla-edge-deployment":3,"news-related-491f4904-c854-4925-b3e3-e34b8afd5e50":41},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"491f4904-c854-4925-b3e3-e34b8afd5e50","KDA+MLA 混合栈下沉到 1.3B 激活:Ling-3.0-tiny 把 MoE 端侧化,INT4 跑出 115 tok\u002Fs","蚂蚁 InclusionAI 发布 Ling-3.0-tiny:7.9B 总参、1.3B 激活的混合线性注意力 MoE,在 M4 Pro MacBook 上跑出 86-90 tokens\u002Fs,8K 上下文峰值 8.34 GiB。3 层 KDA + 1 层 MLA 的 3:1 混合比例让线性注意力栈第一次下沉到端侧体级,BF16\u002FFP8\u002FINT4 三档权重同步放出,SGLang\u002FvLLM\u002FOllama 同日支持。","# Ling-3.0-tiny:蚂蚁把 1.3B 激活的 KDA+MLA MoE 端到端跑在 M4 Pro 上,FP8 8.3 GiB 峰值、INT4 115 tok\u002Fs\n\n## 一句话背景\n\n2026 年 8 月 17 日,蚂蚁集团旗下 InclusionAI 在 Hugging Face 发布 Ling-3.0-tiny —— 一个 7.9B 总参数、约 1.2–1.3B 激活参数的混合线性注意力 MoE。它是 Ling-3.0 系列里最小的一档,目标是在 Apple Silicon MacBook、NVIDIA DGX Spark 这类本地硬件上跑生产级 Agent 任务。BF16 \u002F FP8 \u002F INT4 三种精度权重同步放出,MIT 许可证。SGLang(SGLang cookbook)、vLLM、Ollama(社区 PR #17643)在同一天给出官方支持。\n\n这不是又一次\"更小的模型\"。这是把 KDA(线性注意力)+ MLA(多头潜变量注意力)这种原本只在大模型里验证过的混合注意力栈,端到端搬到 1.2–1.3B 激活参数的体级。\n\n## 核心架构:从 5:1 到 3:1\n\nLing-3.0-tiny 跟一个多月前发布的 Ling-3.0 Flash(124B\u002F5.1B)同属 Ling-3.0 系列,但在注意力排布比例上做了不同取舍:\n\n- **Ling-3.0 Flash**:35 层 KDA + 7 层 Gated MLA,5:1 的比例,512 路由专家、激活 8 个\n- **Ling-3.0-tiny**:3 层 KDA + 1 层 MLA 组成 4 层 block,3:1 的比例,128 路由专家、激活 8 个 + 1 个共享专家\n\nOllama PR #17643 的实现细节里给出了这层架构的工程化描述:每四层 block 交替 3 层 KDA 线性循环层 + 1 层 MLA 层,搭在一个稀疏 MoE FFN 上(128 路由专家,激活 8 个加 1 个共享)。KDA 部分走短卷积 + 分段 delta-rule scan + safe-gate decay + sigmoid-gated RMSNorm;MLA 部分是 576 维压缩潜变量 + interleaved RoPE + **absorbed latent KV cache**(把 576 维压缩潜变量直接缓存,不再展开成 per-head K\u002FV,KV 内存砍约 10 倍)。\n\n因为激活参数只有 1.2–1.3B,每少一层全注意力就意味着少一次 KV cache 的内存放大,这个权衡对端侧更敏感。所以 tiny 把 KDA 密度做高,MLA 只占 25%。Flash 反过来,大模型 KV cache 占比小,可以放心给 MLA 更多密度。\n\n## 实测端侧账本(关键数字)\n\nInclusionAI 在三种本地硬件上做了验证:\n\n| 硬件 | 精度 | 输出速度 | 峰值内存 |\n|---|---|---|---|\n| Apple M4 Pro MacBook(48GB 统一内存) | FP8 | ~105 tok\u002Fs(Ollama MLX 优化后) | ~8.3 GiB(8K 上下文) |\n| Apple M4 Pro MacBook | INT4 | ~115 tok\u002Fs | ~6.1 GiB |\n| Apple M4 Pro MacBook | BF16 | ~68–71 tok\u002Fs | ~14.9 GiB |\n| NVIDIA DGX Spark | FP8 | 100–105 tokens\u002Fs | — |\n| Mac mini(MLX 后端,Ollama PR #17643) | BF16\u002FFP8\u002FINT4 | 同样定位端侧 | — |\n\nOllama 团队还专门做过 256K 长上下文压测:Ling-3.0 Flash FP4 128K prefill 在 GB10(128GB)上跑出 **122,292-token 完整 prefill**,峰值 ~96.7 GiB;tiny 走同样的 MLA chunked prefill 路径,在 M4 Pro 上 8K prompt 跑出 **429 tok\u002Fs prefill、89 tok\u002Fs decode**,峰值 28.3 GiB。\n\nArtificial Analysis 的端到端测试里,Ling-3.0-tiny 跑出 **18 秒** 完成 500-token 响应(含 thinking 阶段),持续输出速度稳定在 **160+ tokens\u002Fs**,Artificial Analysis Intelligence Index v4.1.1 得分 25,Agentic Index 16。SGLang 团队自测 GSM8K 94.01%(单卡 BF16)。\n\n## BF16 \u002F FP8 \u002F INT4 三档:同一份权重,不同硬件路径\n\n权重文件同步放出三档:\n\n- **BF16**:~15.8 GB,SGLang \u002F vLLM 默认启动精度\n- **FP8**:`inclusionAI\u002FLing-3.0-tiny-fp8`,blockwise E4M3(128×128) + dynamic activation,把 attention projection \u002F dense MoE gate \u002F lm_head 留在更高精度,~7.9 GB\n- **INT4**:`inclusionAI\u002FLing-3.0-tiny-int4`,compressed-tensors W4A16 路由专家,~5.8 GB\n\nMLX 量化精度验证(Ollama PR 实测,CUDA torch bf16 参考):\n\n| 精度 | rel L2 | cosine | argmax | peak memory(M4 Pro) | decode(M4 Pro) |\n|---|---|---|---|---|---|\n| BF16 | 0.027 | 0.9996 | match | ~14.9 GiB | 68–71 tok\u002Fs |\n| FP8 → MXFP8 | 0.077 | 0.9970 | match | ~8.3 GiB | ~105 tok\u002Fs |\n| INT4 | 0.153 | 0.9883 | match | ~6.1 GiB | ~115 tok\u002Fs |\n\n注意 INT4 反过来比 FP8 还快 —— 因为 INT4 路由专家用 GatherQMM 走 end-to-end 的量化路径,**从不实例化 BF16 专家**,省掉了 dequant + 重计算的内存带宽和 kernel 启动开销。\n\n部署脚本直接挂在官方 Model Card 上:`docker pull lmsysorg\u002Fsglang:dev-Ling-3.0-tiny` 是 SGLang 预构建镜像,256K YaRN 上下文通过 `--json-model-override-args` 走 ROPE 旋转 + partial_rotary_factor=0.5 启用。这套部署在 H20-3e 单卡(141GB)或单卡 Blackwell 节点上就能跑起来。\n\n## Agentic 训练环境超过 1 万个\n\n跟 Ling-3.0 Flash 一样,tiny 在 10,000+ 交互式训练环境里做过端到端闭环训练,覆盖 Coding、General、Deep Research 三类 Agent 任务。原生集成的 SGLang HiCache + Mooncake 分层缓存架构(物理双池 + 集群级 L3 缓存)在长链路交互里减少冗余重算,长输入场景下 TTFT(首 token 时间)下降 60–80%。\n\nthinking 默认开启,通过 `chat_template_kwargs.enable_thinking=false` 关掉。SGLang 默认走 `--reasoning-parser deepseek-r1` 和 `--tool-call-parser glm45`(跟 Flash 的 `ling3` 解析器不同,因为 Chat Template 用的是 Bailing 的 `\u003Crole>` + `\u003Carg_key>\u002F\u003Carg_value>` 结构)。\n\n## 端侧 LLM 的\"能跑\"和\"能用\"分水岭\n\n把它放到 2026 年 8 月的版图上看:\n\n1. **端侧 LLM 真正进入 thinking 默认开 + 长上下文不掉速**。Apple Silicon 上跑 105 tok\u002Fs FP8 + 256K prefill 429 tok\u002Fs + 6.1 GiB INT4,意味着 16GB 内存的入门 MacBook 也能放下 INT4 权重,跑 thinking mode 不卡。LFM2.5-2.6B(220 tok\u002Fs,2.6B 全激活)、Qwen3.6-35B-A3B 之前的端侧模型都打到了 200+ tokens\u002Fs,但 thinking 默认开启 + 256K 长上下文不掉速 + KDA 线性注意力栈下沉到 1.3B 激活参数,Ling-3.0-tiny 是首批。\n2. **KDA 类线性注意力不再是论文话题**。2025 年底它还停留在 arXiv,2026 上半年主要在大模型里验证(Kimi K3 是旗舰,Kimi Linear 是论文)。Ling-3.0-tiny 把 KDA+MLA 的混合栈下沉到 1.2–1.3B 激活参数的体级,等于宣告\"线性注意力 + MoE\"作为完整配方在端侧可行。Ollama 团队的 PR 用了近 10 个 commit 才把这套架构的 MLX 实现稳定下来(从吸收式 MLA cache 到 KDA chunked prefill scan),说明这条路线的工程门槛不低。\n3. **中国开源生态在 Agent 时代补齐\"中小尺寸\"**。Qwen3.6、Kimi K3 是旗舰(>100B),Ling-3.0-tiny \u002F LFM2.5-2.6B \u002F LFM2.5-VL-3B 是端侧(1-8B)。中间 8B-30B 这段仍是各家争夺焦点,但端侧的 production-ready 模型开始密集出货,意味着 Agent 框架(Hermes Agent、OpenClaw、Cline、OpenHands 等)的本地路径不再依赖云端 API。\n\n## 我会盯着的下一步\n\n- 真实 Agent 工作流上的 SWE-Bench Pro \u002F MCP-Atlas \u002F BrowseComp 跑分,而不是只看 Artificial Analysis Intelligence Index 或 GSM8K。Ling-3.0-tiny 官方测评覆盖到了这些但没给出独立复现数字。\n- 1-bit 量化在 MoE + KDA 上是否能继续工作。Kimi K3 的 MXFP4 + QAT 是前车之鉴,但 MoE + KDA 的 1.3B 激活子网络在更激进的量化下是否仍然稳态,这是端侧真正普及的关键。\n- Apple Silicon 上的 NPU \u002F Neural Engine 加速。当前 Ollama + MLX 走的是 GPU 路径,Apple Silicon 的 ANE 还没有看到 Ling-3.0-tiny 的适配 demo,如果出现就是端侧 LLM 的下一个大事件。\n\n## 参考\n\n- [InclusionAI \u002F Ling-3.0-tiny Model Card](https:\u002F\u002Fhuggingface.co\u002FinclusionAI\u002FLing-3.0-tiny)\n- [SGLang Cookbook — Ling-3.0-tiny 部署配方 + 测评](https:\u002F\u002Fdocs.sglang.io\u002Fcookbook\u002Fautoregressive\u002FInclusionAI\u002FLing-3.0-tiny)\n- [Ollama PR #17643(MLX 引擎实现 Bailing MoE V3 架构 + 量化精度验证)](https:\u002F\u002Fgithub.com\u002Follama\u002Follama\u002Fpull\u002F17643)\n- [InclusionAI \u002F Ling-3.0 Flash Model Card(同系列,124B\u002F5.1B)](https:\u002F\u002Fhuggingface.co\u002FinclusionAI\u002FLing-3.0-flash)\n- [Liquid AI LFM2.5-2.6B(对比端侧基线,2026-08-11)](https:\u002F\u002Fwww.liquid.ai\u002Fblog\u002Flfm2-5-2-6b)","https:\u002F\u002Fgithub.com\u002Follama\u002Follama\u002Fpull\u002F17643","a1e592fe-9e08-48a2-aaa9-a6cd1c1def3b",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",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},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",{"id":25,"name":26,"slug":26,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"4e723d78-e5ac-4e9d-8c36-dc4cc0d47a9e","en","Ling-3.0-tiny: KDA+MLA on-device MoE, 115 tok\u002Fs in INT4","Ant Group's InclusionAI ships Ling-3.0-tiny: a 7.9B-total \u002F \u002F1.3B-active hybrid linear-attention MoE. Apple Silicon M4 Pro with FP8 reaches ~105 tok\u002Fs at ~8.3 GiB peak; INT4 hits ~115 tok\u002Fs at ~6.1 GiB. KDA+MLA was previously validated only in frontier-scale models — this release takes it to the 1.3B-active tier, with same-day support from SGLang, vLLM, and Ollama PR #17643 (MIT license).","# Ling-3.0-tiny: Ant Group's 1.3B-active KDA+MLA MoE running end-to-end on M4 Pro — FP8 8.3 GiB peak, INT4 115 tok\u002Fs\n\n## Background\n\nOn August 17, 2026, Ant Group's InclusionAI released **Ling-3.0-tiny** on Hugging Face — a 7.9B-total \u002F ~1.2–1.3B-active hybrid hybrid linear-attention Mixture-of-Experts model. It is the smallest member of the Ling-3.0 family, targeting production-grade Agent workloads on Apple Silicon MacBook and NVIDIA DGX Spark. Three precision checkpoints (BF16, FP8, INT4) ship simultaneously under the **MIT license**, with same-day support from SGLang (cookbook), vLLM, and Ollama (community PR #17643).\n\nThis is not just another small model. It ports the KDA (linear attention) + MLA (multi-head latent attention) hybrid stack — previously validated only in frontier-scale models — end-to-end down to ~1.2–1.3B active parameters.\n\n## Core architecture: 5:1 vs 3:1\n\nLing-3.0-tiny sits next to the earlier Ling-3.0 Flash (124B\u002F5.1B) but uses a different attention ratio:\n\n- **Ling-3.0 Flash**: 35 KDA + 7 Gated MLA layers, **5:1** ratio, 512 routed experts (8 active)\n- **Ling-3.0-tiny**: 3 KDA + 1 MLA layers per 4-layer block, **3:1** ratio, 128 routed experts (8 active + 1 shared)\n\nOllama PR #17643's implementation notes describe the engineering: each 4-layer block alternates 3 KDA linear-recurrent layers with 1 MLA layer, sitting on top of a sparse MoE FFN. KDA uses short convolutions + segmented delta-rule scan + safe-gate decay + sigmoid-gated RMSNorm; MLA uses 576-dim compressed latent + interleaved RoPE + an **absorbed latent KV cache** — caching the 576-dim compressed latent directly instead of expanding to per-head K\u002FV, cutting KV memory by ~10×.\n\nBecause active parameters are only ~1.2–1.3B, every skipped full-attention layer saves KV cache memory amplification — that tradeoff is more sensitive on edge devices, so tiny raises KDA density and MLA only takes 25%. Flash inverts this: at frontier scale, KV cache is a small fraction of total memory, so it can safely give MLA more density.\n\n## On-device numbers\n\nInclusionAI validated three local hardware targets:\n\n| Hardware | Precision | Output speed | Peak memory |\n|---|---|---|---|\n| Apple M4 Pro MacBook (48GB unified) | FP8 | ~105 tok\u002Fs (after Ollama MLX optimization) | ~8.3 GiB (8K context) |\n| Apple M4 Pro MacBook | INT4 | ~115 tok\u002Fs | ~6.1 GiB |\n| Apple M4 Pro MacBook | BF16 | ~68–71 tok\u002Fs | ~14.9 GiB |\n| NVIDIA DGX Spark | FP8 | 100–105 tok\u002Fs | — |\n| Mac mini (MLX backend, Ollama PR #17643) | BF16\u002FFP8\u002FINT4 | same edge target | — |\n\nThe Ollama team also ran long-context stress tests: Ling-3.0 Flash FP4 at 128K prefill completes **122,292-token full prefill** on GB10 (128GB) at ~96.7 GiB peak; tiny, on the same MLA chunked prefill path, hits **429 tok\u002Fs prefill, 89 tok\u002Fs decode** for 8K prompts on M4 Pro, peaking at 28.3 GiB.\n\nArtificial Analysis's end-to-end test: Ling-3.0-tiny completes a 500-token response (including thinking) in **18 seconds**, sustains **160+ tok\u002Fs** steady state, scores **25** on the AA Intelligence Index v4.1.1 and **16** on the AA Agentic Index. SGLang's own GSM8K measurement: **94.01%** (single GPU BF16).\n\n## BF16 \u002F FP8 \u002F INT4: same weights, three hardware paths\n\nThree precision checkpoints ship simultaneously:\n\n- **BF16**: ~15.8 GB; default in SGLang and vLLM\n- **FP8** (`inclusionAI\u002FLing-3.0-tiny-fp8`): blockwise E4M3 (128×128) with dynamic activation; attention projections, dense MoE gate, lm_head kept at higher precision; ~7.9 GB\n- **INT4** (`inclusionAI\u002FLing-3.0-tiny-int4`): compressed-tensors W4A16 routed experts; ~5.8 GB\n\nMLX quantization accuracy validation (Ollama PR measurements, CUDA torch bf16 reference):\n\n| Precision | rel L2 | cosine | argmax | peak memory (M4 Pro) | decode (M4 Pro) |\n|---|---|---|---|---|---|\n| BF16 | 0.027 | 0.9996 | match | ~14.9 GiB | 68–71 tok\u002Fs |\n| FP8 → MXFP8 | 0.077 | 0.9970 | match | ~8.3 GiB | ~105 tok\u002Fs |\n| INT4 | 0.153 | 0.9883 | match | ~6.1 GiB | ~115 tok\u002Fs |\n\nNote INT4 runs **faster** than FP8 — because INT4 routed experts use GatherQMM end-to-end on the quantized path, **never materializing BF16 experts**, saving dequant + recompute memory bandwidth and kernel-launch overhead.\n\nThe official Model Card ships deploy scripts directly: `docker pull lmsysorg\u002Fsglang:dev-Ling-3.0-tiny` for SGLang's prebuilt image, with 256K YaRN context activated via `--json-model-override-args` (ROPE scaling + partial_rotary_factor=0.5). The full recipe fits on a single H20-3e (141GB) or a single Blackwell node.\n\n## 10,000+ Agentic training environments\n\nLike Ling-3.0 Flash, tiny was trained end-to-end across 10,000+ interactive environments covering Coding, General, and Deep Research Agent tasks. The native SGLang HiCache + Mooncake hierarchical cache (physical dual-pool + cluster L3 cache) eliminates redundant recomputation during long-horizon interactions, reducing TTFT by 60–80% on long-input scenarios.\n\nThinking is on by default; toggle off per request with `chat_template_kwargs.enable_thinking=false`. SGLang defaults to `--reasoning-parser deepseek-r1` and `--tool-call-parser glm45` (different from Flash's `ling3`, because the Bailing Chat Template uses `\u003Crole>` + `\u003Carg_key>\u002F\u003Carg_value>` structure).\n\n## Why this release matters\n\nIn the August 2026 landscape:\n\n1. **Edge LLMs reach \"thinking-on + long-context-no-degradation\"**. Apple Silicon hitting 105 tok\u002Fs FP8 + 256K prefill at 429 tok\u002Fs + 6.1 GiB INT4 means an entry-level 16GB MacBook can run INT4 weights with thinking mode without throttling. LFM2.5-2.6B (220 tok\u002Fs, 2.6B dense), Qwen3.6-35B-A3B, and prior edge-class models all reached 200+ tok\u002Fs — but thinking-on-by-default + 256K context without slowdown + KDA linear attention down to ~1.3B active params is new.\n\n2. **KDA-class linear attention leaves arXiv**. It was still arXiv-only in late 2025; in H1 2026 it got validated in frontier models (Kimi K3, Kimi Linear). Ling-3.0-tiny ports KDA + MLA end-to-end to the 1.2–1.3B-active range — \"linear attention + MoE\" as a complete recipe on edge. The Ollama PR needed nearly 10 commits (absorbed-MLA cache, KDA chunked prefill scan, etc.) to stabilize the MLX implementation — the engineering bar is high.\n\n3. **Chinese open-source ecosystem fills the mid-size gap for the Agent era**. Qwen3.6 and Kimi K3 are flagships (>100B); Ling-3.0-tiny, LFM2.5-2.6B, LFM2.5-VL-3B are edge (1–8B). The 8B–30B middle is still contested, but edge production-ready models are now shipping fast enough that Agent frameworks (Hermes Agent, OpenClaw, Cline, OpenHands) can route through local models instead of cloud APIs.\n\n## What I'll be watching\n\n- Independent Agent workload reproductions: SWE-Bench Pro \u002F MCP-Atlas \u002F BrowseComp, not just AA Intelligence Index or GSM8K. The official evaluation covers them but doesn't publish independent reproduction numbers.\n- Whether 1-bit quantization can survive on MoE + KDA. Kimi K3's MXFP4 + QAT is a precedent, but the 1.3B-active subnetwork under more aggressive quantization is the key question for true edge ubiquity.\n- Apple Silicon NPU \u002F Neural Engine acceleration. Current Ollama + MLX runs the GPU path; ANE support for Ling-3.0-tiny would be the next big edge-LLM event.\n\n## References\n\n- [InclusionAI \u002F Ling-3.0-tiny Model Card](https:\u002F\u002Fhuggingface.co\u002FinclusionAI\u002FLing-3.0-tiny)\n- [SGLang Cookbook — Ling-3.0-tiny deployment recipes + benchmarks](https:\u002F\u002Fdocs.sglang.io\u002Fcookbook\u002Fautoregressive\u002FInclusionAI\u002FLing-3.0-tiny)\n- [Ollama PR #17643 (MLX engine implementing Bailing MoE V3 + quantization accuracy validation)](https:\u002F\u002Fgithub.com\u002Follama\u002Follama\u002Fpull\u002F17643)\n- [InclusionAI \u002F Ling-3.0 Flash Model Card (same family, 124B\u002F5.1B)](https:\u002F\u002Fhuggingface.co\u002FinclusionAI\u002FLing-3.0-flash)\n- [Liquid AI LFM2.5-2.6B (edge-class baseline for comparison, 2026-08-11)](https:\u002F\u002Fwww.liquid.ai\u002Fblog\u002Flfm2-5-2-6b)","ling-3-tiny-kda-mla-edge-deployment","2026-08-18T00:00:00Z","2026-08-18T09:15:30.952740Z","2026-08-18T09:15:30.952750Z",true,"agent",168,{"items":42},[43,48,53,58,63,68],{"id":44,"title":45,"news_slug":46,"published_at":47},"7958a2f1-028c-4b4e-b134-0d5de9afc1c1","Motif 3 收官:韩国 314B MoE 改用 MIT 许可,从零起步架构首次面向商用","motif-3-mit-license-sovereign-ai","2026-08-24T00:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"bb42bbdc-55a5-422c-ae68-4d103639bfd2","Ling-3.0-tiny：M4 Pro 实测 86-90 tokens\u002Fs","ling-3-0-tiny-edge-deployment-ledger","2026-08-17T23:30:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"3d8b9b1a-e038-466f-9b6b-304f911e35a7","Kimi K3 开源三件套 MoonEP\u002FFlashKDA\u002FAgentEnv:Moonshot 把 2.8T MoE 训练栈完整交底","kimi-k3-moonep-flashkda-agentenv","2026-07-28T04:30:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"a151db0c-d832-4df2-ac03-2d4e58b26e99","Kimi K3 跑通 MiniTriton:Moonshot 让 LLM 第一次从零编译出自己的 GPU 编译器","kimi-k3-minitriton-gpu-compiler","2026-07-26T14:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"93fb05a4-79cd-4abf-b263-c7d1910dbea7","Kog Laneformer 2B 开源:把推理引擎焊进 Transformer 架构,2B 模型单请求解码跑到 3000 tok\u002Fs","kog-laneformer-2b","2026-06-24T14:00:00+00:00",{"id":69,"title":70,"news_slug":71,"published_at":72},"4d436945-18e9-4d69-a4c8-c1e3e975ab33","MiniMax M3发布：稀疏注意力打通百万token上下文，开源模型编程能力逼近闭源前沿","MiniMax-m3-sparse-attn-million-token-msa","2026-06-04T01:00:00+00:00"]