vLLM Delivers Day-0 Support for Kimi K3.1 Pro
Kimi K3.1 Pro is a 2.8T-parameter (104B active) multimodal MoE with 1M context; vLLM ran it on day 0 across NVIDIA, AMD, TPU and T-Head hardware.
AI Weekly Recap
35 articles · 5 categories
Weekly pattern report
5 shifts that shaped AI this week
2026-10-31 → 2026-11-06
2026-W45 · 35 articles reviewed
The week in signals
- Kimi K3.1 Pro (2.8T MoE, 1M context) shipped open weights with day-0 vLLM and Ascend support.
- Same model, different agent: 82.0% vs 65.4% across 21 MCP servers.
- A planner/coder/reviewer team passed 25.7% of tests versus 56.8% for one Codex agent.
- Mandiant: stolen LLM API keys produced nearly $1M in charges at one company.
- An H200 loses money even fully busy; a B300 profits at 35% utilization.
Open-weight Kimi K3.1 Pro landed with day-0 serving from vLLM, Huawei Ascend and hosted providers, so launch-day infrastructure support is now part of a model release.
On the agent side, measurements said the harness matters more than the model: the same Haiku scored 82.0% or 65.4% depending on the agent, and a 17-agent setup lost to vanilla Claude Code.
Security followed: MCP servers, coding-tool browsers and stolen API keys were all exploited. Build controls around the tool call, not the prompt.
Prefer to watch?
LLM Digest Weekly Explainer: Oct 31-Nov 6, 2026
Kimi K3.1 Pro and the Open-Weight Serving Race 6 items
Moonshot's open-weight K3.1 Pro shipped with day-0 support across vLLM, Huawei Ascend and hosted inference, so serving stacks now compete on launch-day readiness.
Moonshot AI Publishes Kimi K3.1 Pro Weights on Hugging Face
Moonshot published the K3.1 Pro weights on Hugging Face.
Huawei Ascend Adds Day-Zero Support for Kimi K3.1 Pro Across Atlas A3 and Ascend 950 SuperPoDs
Huawei Ascend added day-zero K3.1 Pro support on Atlas A3 and Ascend 950 SuperPoDs.
Verda Inference, in Preview with Kimi K3.1 Pro
Verda opened a preview of serverless Kimi K3.1 Pro inference.
Moonshot AI Launches Hosted Agent Service, Sandbox Priced at 0.072 Yuan per Hour
Moonshot launched a hosted agent service, with sandboxes priced at 0.072 yuan per hour.
[AINews] Inkling 2 Small: 17A-594B trained in 68 hours on 12T tokens, beating Muse Spark
Latent Space's roundup covers Inkling 2 Small (17A-594B, trained in 68 hours on 12T tokens) and METR's study of AI research.
Coding-Agent Harness Benchmarks 8 items
This week's measurements show the harness and hand-offs, not the model, often decide agent results.
Same model, different agent: 17 points apart on 21 MCP servers
Claude Haiku 5.5 answered 82.0% of tasks over 21 MCP servers via OpenCode but 65.4% via Claude Code; built-in tools explain the gap.
We Tested Coding Agent Harnesses. Thinking Harder Beat All of Them
On 100 SWE tasks, a setup with 37 skills and 17 agents cost 8x more and scored below vanilla Claude Code; higher thinking effort beat all four setups.
Does a team of coding agents beat one agent?
Codex alone passed 56.8% of hidden tests; a planner/coder/reviewer team passed 25.7% because the planner's summary lost details.
Nvidia Study: AI Agents Rarely Need Long Context Windows
NVIDIA's preprint finds Claude Opus 4.5 and GPT 5.6 Sol lose little accuracy at 16K context with frequent compaction, while long-context serving costs up to 7x more per token.
Which coding-agent CLI is best for headless workflow automation?
Six headless coding-agent CLIs compared on start-up, time per turn, memory, image size and tokens; pi had the lowest overhead.
Harness 2: executing agent code as it streams
Framer's rebuilt harness lets the agent write JavaScript that runs as it streams, making workflows up to 33% cheaper and 20% faster.
Agents Are Going Async
LangChain argues agents must stay responsive while working on long tasks in the background.
How We Optimized Browser Use for Deep Agents
LangChain details how it optimized browser use for Deep Agents.
Agent Security Incidents and Guardrails 8 items
Attackers hit agent tooling and API keys directly this week, and builders responded with scoped identity and call-level controls.
How the built-in browsers in five AI coding tools broke basic browser rules
Forever Security's EMBRACE research showed built-in browsers in five AI coding tools could run commands, read the disk or spoof the address bar.
AI coding agents are installing abandoned MCP servers with known security issues
Glow Labs found coding agents installing deprecated MCP servers with known high-severity vulnerabilities.
Stolen AI API keys generated nearly $1M in charges for a single company
Mandiant reports LLMjacking intrusions in 2026, including stolen API keys that generated nearly $1M in charges at one company.
How We Took Over ChatGPT Accounts
Parameter Labs describe an exploit against ChatGPT's artifacts implementation that allowed account takeover.
You Had Me at "Hello": A Chinese‑Language Framework Targeting AI Infrastructure
Era researchers traced a Chinese-language framework targeting AI infrastructure; logs held AWS credentials and cloud metadata probe results.
Show HN: Agentic Identity Broker, OAuth Delegation for AI Agents
Zalando's MIT-licensed Agentic Identity Broker gives agents scoped, revocable OAuth access to MCP services.
Show HN: Replay three AI agent approval failures in five seconds, in Python
Call4Agents makes Python agent actions safe to retry so approved tool calls run at most once.
Show HN: MCPgraphy – a proxy that learns when your agent's MCP tools work
MCPgraphy is an MCP proxy that learns which tools work in which state and warns or blocks the next bad call.
Evals and Agent Reliability 7 items
New benchmarks and eval studies show agent scores overstate real-world performance.
GDPx agent benchmark: cost per point for 12 models, and the safety-filter caveat
SemiAnalysis's GDPx has 206 long agent tasks across 11 fields; the best model scored 12%.
Eval-production behavior gaps in agent safety and the role of eval-awareness
A study comparing agent behavior in safety evals versus deployment finds eval-awareness opens a gap that is now being tracked directly.
Is AI Cheap Enough to Replace Developers?
Claude Opus 4.8 clears 88.6% on SWE-bench Verified but drops sharply on the private-codebase SWE-bench Pro.
How we made evals 100× cheaper with multi-agent architecture
Cleric rearchitected its eval agent into about 40 loosely coupled agents, running evals 5x faster and ~100x cheaper.
We tested 12 search APIs for AI agents. Even the best one isn't enough
Twelve search APIs for agents tested; combining providers lifted accuracy to 82.6% with four, versus 75.5% with two.
Claude Platform now supports native OpenTelemetry observability in Databricks MLflow
Claude Platform now exports native OpenTelemetry traces to Databricks MLflow for finding recurring agent failures.
LangSmith BYOC is now available on AWS, Azure, and Google Cloud
LangSmith BYOC is GA on AWS and Beta on Azure and Google Cloud, keeping traces and datasets in the customer's account.
Inference Infrastructure and Economics 6 items
Control planes, new runtimes and unit economics all point to GPU choice and orchestration outweighing engine choice.
Introducing Kinference: An Open Source Inference Control Plane for vLLM and Kubernetes
Buoyant open-sourced Kinference, a control plane for vLLM on Kubernetes.
Distributed Inference Benchmarks: [K]Inference vs. Plain Kubernetes
Kinference benchmarks compare it against plain Kubernetes on heterogeneous GPUs.
Supercharging XLA-Based Inference: Announcing the Launch of XLA Serving Engine
Google open-sourced XLA Serving Engine for a train-in-Python, serve-in-C++ workflow.
Is hosting an open-source inference engine profitable?
An analysis finds an H200 loses money even fully busy, while a B300 profits once 35% busy.
Go vs. Python for an LLM Gateway: What We Measured
A benchmark of the Go gateway KTMS against LiteLLM Proxy and eight other self-hosted gateways on latency, throughput and memory.
The Iron(wood) Throne: Google's TPU Stakes Its Claim on Inference
SemiAnalysis argues Google's TPUs are gaining inference share, with Oracle, OpenAI, Anthropic and Meta as customers.
The week, resolved into patterns