A project shaped by experimentation, trust and a genuinely wonderful collaboration. Thank you to the Kimi team for believing in the process and being such a pleasure to create with! #KIMI# #KIMIK3# #MOONSHOTAI# #K3# #BTS#

🧩 Kimi K3’s MoE and Attention Are Built Around Trade-offs, Not Tricks Kimi K3’s open release has drawn attention to its scale. But its architecture tells a more useful story: the hardest part of scaling is keeping quality, efficiency, and stability in balance. Zhihu contributor 苏剑林 @Jianlin_S explains the design logic behind two core components: Stable LatentMoE and K3’s hybrid attention. At a high level: K3 = KDA + MLA + Stable LatentMoE + AttnRes 1️⃣ Stable LatentMoE: more experts at similar cost LatentMoE compresses each token into a smaller latent space before routing it to experts, then projects the result back to the full hidden dimension. This reduces expert computation and communication. The saved budget can support more, narrower experts without greatly increasing training or inference cost. But the longer projection chain also magnifies numerical instability. K3 introduces three fixes. 🔹 SiTU-GLU softly caps extreme activations in both branches of the expert network. Compared with hard clipping, soft capping preserves smoother optimization. 🔹 RMSNorm is placed before the final up-projection. It stabilizes training and helps balance routed experts against shared experts. 🔹 Quantile Balancing replaces the previous load-balancing update, which became unreliable as the expert pool grew. It approximates global routing quantiles with histograms, allowing efficient aggregation across machines. The broader lesson is clear: scaling MoE is not just about adding experts. Routing, activation ranges, normalization, and distributed communication must scale with them. 2️⃣ Why K3 still uses MLA Some newer models have moved away from MLA, partly because speculative decoding changes the inference trade-off. MLA keeps KV Cache small and remains highly competitive under fixed training and memory budgets. But its decoding path is relatively compute-heavy, leaving less room for Multi-Token Prediction to trade extra computation for speed. Other attention designs simply move the bottleneck: 🔹 Smaller designs may reduce computation but lose quality or require a larger KV Cache. 🔹 Larger designs can recover quality, but increase training and prefill costs. An ideal replacement would preserve quality, reduce KV Cache, lower decoding compute, and cost no more during training or prefill. No simple design currently satisfies all four conditions. K3 therefore keeps MLA and combines it with KDA. The linear-attention layers handle most long-context processing efficiently, while MLA preserves full-attention capacity where it matters. 3️⃣ “Abandoning MLA” is not so simple Architectures that appear to replace MLA may still retain its core intuition. For example, a wide MQA design with shared K and V resembles MLA’s decoding form. Sparsity and compression can then reduce its compute and cache costs. This can work, but it introduces more infrastructure complexity. So the current debate is less about whether MLA is obsolete. It is about which combination of full, linear, sparse, and compressed attention offers the best system-level trade-off. 4️⃣ Why K3 can remove RoPE K3 removes RoPE from its MLA layers. That would hurt a pure-MLA model. But K3 is a hybrid of KDA and MLA. KDA’s DeltaNet-style updates already introduce an implicit positional transformation. In this sense, KDA provides something similar to a generalized form of RoPE for the full network. Adding explicit RoPE back produced little difference, so K3 followed the simpler design. K3 is not truly position-free. Its positional structure is partly carried by KDA instead of an explicit embedding. ⚙ The real architecture lesson None of these choices is especially flashy in isolation. Stable LatentMoE controls the numerical and routing problems created by more experts. KDA and MLA divide long-context work according to their strengths. NoPE removes a redundant component only after the hybrid architecture makes it unnecessary. K3’s main design principle is therefore not novelty for its own sake. Every architectural change must justify itself across quality, efficiency, and stability. 🔗 Full reading: 📖Blog post: #KimiK3# #MoE# #Attention# #LLM# #AIInfra# #OpenSourceAI#

ZERO Market Flash #003# China's Second AI Shockwave: Is the Market Misreading It Again? Why Cheaper AI May Mean More Infrastructure—Not Less In January 2025, DeepSeek triggered one of the biggest debates in AI investing. The market quickly concluded that if powerful models could be trained with fewer GPUs, future demand for AI infrastructure must decline. NVIDIA lost nearly 17% in a single day, and AI-related stocks sold off across the board. More than a year later, history appears to be repeating itself. Kimi K3 has once again demonstrated that Chinese companies can build highly competitive large language models at significantly lower cost. The market immediately returned to the same question: If AI keeps getting cheaper, will we need fewer GPUs? Ironically, Kimi itself may have provided the opposite answer. Shortly after launch, the company suspended new subscriptions—not because the model had reached its limits, but because user demand had pushed GPU capacity close to its deployment limit. That may be the most important signal from Kimi's release. The first bottleneck wasn't model capability. It was deployment capacity. For the past several years, AI competition has largely been defined by training. Whoever trained larger models with more GPUs was assumed to have the strongest competitive advantage. Under that framework, lower training costs naturally imply lower infrastructure demand. But that assumption depends on one premise: that AI's value is created primarily during training. The more important question is: What happens if cheaper AI leads to dramatically more adoption? A foundation model may be trained once. It may perform billions of inference requests afterward. Over the long run, infrastructure consumption is driven less by training than by continuous deployment. Lower cost reduces the price of each interaction. Growing adoption increases the number of interactions. If usage grows faster than cost declines, total infrastructure demand can continue to expand. That is why Kimi's GPU capacity announcement may matter more than the model itself. The market focused on lower training costs. Reality exposed growing deployment demand. This also gives new context to SK Group Chairman Chey Tae-won's observation that the memory industry may gradually shift from a Price-driven cycle to a Volume-driven one. If AI deployment continues to expand, future industry growth may depend less on rising prices and more on rising deployment volumes. Kimi's capacity constraints do not prove that transition has already happened. They do suggest that AI competition is beginning to extend beyond training and into deployment. One year ago, DeepSeek forced investors to rethink training costs. Today, Kimi may be forcing investors to rethink deployment demand. Training creates models. Deployment creates industries. — This article reflects personal research and opinions only and should not be considered investment advice. Please conduct your own research before making investment decisions. ZERO Good is not good enough for conviction.
Only the best deserves concentration. Scientist · Doctor · A9 Investor Search Tags #AI# #ArtificialIntelligence# #GenerativeAI# #LLM# #KimiK3# #MoonshotAI# #DeepSeek# #Inference# #Deployment# #Training# #AIAgents# #GPU# #NVIDIA# #Memory# #HBM# #Semiconductors# #AIInfrastructure# #SKHynix# #Micron# #SNDK# #TechInvesting# #ZERO#

🧩 Your Agent Model May Be Overfitting the Harness, Not Learning the Task DeepSeek V4 Pro has exposed a growing Agent problem: the same weights can approach their ceiling under DSH’s minimal preset, then degrade under standard or third-party frameworks. Zhihu contributor 曾天真 compares reports from Kimi K3, Qwen, Kwai, and DeepSeek, then connects them to his team’s production experience. The core lesson: a model can understand the task while remaining unable to execute it outside its training interface. 1️⃣ What is harness overfitting? An Agent harness defines system prompts, tool schemas, history layout, result truncation, planning hooks, reflection timing, and stopping rules. Kwai’s KAT-Coder-V2.5 report divides overfitting into three types: 🔹 Format overfitting: changing the tool-call protocol causes parsing failures. 🔹 Context-structure overfitting: rearranging history, truncating results, or enabling compaction changes behavior. 🔹 Control-flow overfitting: planning and stopping depend on scaffolding provided by the training harness. The last type is especially dangerous. If a model stops planning because the new runtime has no todo tool, it has learned a protocol rather than a transferable capability. 2️⃣ Kimi K3: test with an unfamiliar harness The Kimi K3 report treats diverse, verifiable environments as a prerequisite for Agent RL. Its results explicitly name the harness used. Kimi also reserves MIRA as an out-of-distribution harness and evaluates every model under the same environment. K3 even reports the same benchmark under Kimi Code and Claude Code, with only a 0.8-point difference. This is stronger evidence of cross-harness stability than a general robustness claim. But K3 still has protocol coupling. It was trained with preserved thinking history. If a harness does not return the complete reasoning history across turns, generation can become unstable. Kimi therefore offers useful methodology, not immunity: task-level environment diversity cannot remove a hard dependency at the message-protocol level. 3️⃣ Qwen provides the cleanest controlled evidence The Qwen3-Coder-Next report states the problem directly: training with one tool-chat template often makes models memorize a particular output structure. Qwen trains across natural-language descriptions, JSON, Python-style calls, XML schemas, and TypeScript interfaces. Its strongest evidence is a controlled ablation. With data volume and training recipe fixed, increasing the number of tool templates improved SWE-bench Verified. Interface diversity may therefore improve the main benchmark, not merely reduce deployment failures. Qwen also evaluates models across five real CLI and IDE scaffolds. During RL, malformed tool calls receive token-level penalties. Qwen3.8 makes reasoning depth and thinking-history preservation configurable. Kimi treats preserved thinking as a requirement; Qwen exposes it as an option. 4️⃣ Harness Scaling must cover the right dimensions Kwai describes its solution as Harness Scaling, or domain randomization applied to Agent rollouts. The key is not the number of harnesses. It is whether they vary along dimensions that matter: 🔹 Tool protocols: structured function calls, code blocks, or tag-based formats. 🔹 Context management: full history, sliding windows, summaries, compaction, and different truncation policies. 🔹 Control flow: minimal ReAct loops versus explicit planning and self-reflection. This broader design matters because tool-format diversity alone cannot address context and control-flow dependence. Kwai also finds that a model may perform better under a simpler harness. More tools can increase unnecessary exploration and weaken stopping behavior. More scaffolding does not always produce a stronger Agent. 5️⃣ DeepSeek’s transparency made its coupling measurable The DeepSeek-V4 report publishes its XML tool-call schema and RL system prompt. In the open-source DeepSeek Harness, minimal keeps only Bash and str_replace_editor, disables context compression, and reproduces the training interface. A snapshot test is explicitly named: “sends the exact RL prompt and schemas” So minimal is not simply a lighter standard preset. It is a reconstruction of the interface used during RL. DeepSeek’s post-training pipeline also raises a broader concern. Domain specialists are trained with specialized prompts and rewards, then merged through On-Policy Distillation. Interface habits learned by those specialists may be distilled alongside genuine capabilities. DeepSeek also preserves complete reasoning history during tool use. Its report warns that frameworks simulating tools through user messages may not activate the intended context path. As with K3, reasoning-history structure becomes an implicit contract between the model and harness. 6️⃣ Production failures reveal what benchmarks miss The author’s team initially used one internal runtime for RL because it was stable, observable, and easy to connect to rewards. Deployment exposed the hidden coupling: · Unfamiliar tool names pushed the model toward shell workarounds. · Truncated tool results caused it to abandon partially correct work. · Without a planning tool, explicit planning disappeared. · Adding more MCP tools increased exploration and weakened stopping. The model still understood the task. It had learned to solve it inside one runtime. The team replaced that runtime with a randomized family of environments. They varied tool names, parameter styles, tool count, result truncation, and context policies, while keeping one canonical configuration for regression testing. This required more environment engineering and slowed debugging and convergence. But the benefit appeared in the worst integration. Average performance barely changed, while variance narrowed and complaints decreased. Teams should therefore track the worst integration score or the performance range across harnesses, not only the mean. 7️⃣ Distillation can carry interface pollution Teacher trajectories contain tool preferences, fixed call sequences, confirmation phrases, and output conventions from the teacher’s harness. Students may learn these artifacts as mandatory behavior, then request nonexistent tools or repeat unsupported boilerplate in production. Three fixes worked best: 🔹 Label task-related and interface-related trajectory segments, then rewrite or mask the latter. 🔹 Generate the same task under multiple harnesses and mix the resulting trajectories. 🔹 Add explicit examples for missing tools, failed calls, truncated results, and incompatible schemas. Finally, the harness must be treated as a versioned dependency. Its version belongs in experiment metadata. Prompt, schema, and truncation changes need review. Models and harnesses should ship with a compatibility matrix. The harness is no longer just infrastructure around the model. It is part of the training distribution and part of the model’s behavior. 🔗 Full analysis: #AIAgents# #AgentHarness# #DeepSeek# #KimiK3# #Qwen# #ReinforcementLearning# #LLM#

Kimi K3 is now on Amazon Bedrock! Run coding, document analysis, and extended agent workflows with Bedrock's access, encryption, and auditing controls. Explicit prompt caching supported. Start building with K3 on AWS 👉

Kimi K3 serving in vLLM now delivers 2.2–2.8x throughput on our B300 benchmark vs v0.27.1. We break down the work across scheduling, KDA state handling, and MoE kernels, with benchmarks and commands to reproduce the results. Thanks to the vLLM community for pushing Kimi K3 performance forward! Read the deep dive: