Moonshot AI's Kimi K2.7 Code Model Now Available on Microsoft Azure Foundry

·1 views

Microsoft and Moonshot AI jointly launched Kimi K2.7 Code on the Azure Foundry model catalog on July 1, 2026, targeting complex, multi-step software engineering workflows. The model features a 256K context window and is optimized for tasks such as large-scale refactoring, cross-service feature implementation, and multi-step debugging sessions. Compared to its predecessor K2.6, K2.7 Code delivers approximately 30% fewer thinking tokens, improving speed and cost efficiency. On the human-verified MCP Mark benchmark, K2.7 Code scored 81.1, outperforming Claude Opus 4.8's 76.4, though it trails GPT-5.5 on pure coding benchmarks. The model is priced at $0.95 per million input tokens and $4.00 per million output tokens, positioning it below GPT-4o-class pricing for agentic and developer tooling use cases.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

How a Marketplace Buy Box Works and Why Getting It Right Is Hard

A marketplace buy box is the mechanism that selects one winning seller offer to appear on the main 'add to cart' button when multiple sellers list the same product. The winning offer is typically chosen through a sequential set of criteria: stock availability comes first, followed by the lowest price, then seller quality signals such as premium-seller status, and finally a fixed tiebreaker rule. Because most buyers purchase without comparing all available offers, holding the buy box position captures the majority of sales for a given product, making it the most commercially valuable spot on a product page. The selection logic must be fully deterministic — returning the same winner on every page load — since any randomness erodes trust for both buyers and sellers. Poorly designed buy box logic can inadvertently reward the wrong seller behaviour, such as favouring the cheapest seller who is chronically out of stock, while sound logic can push sellers toward better pricing, reliability, and service quality.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

New Metrics CD and RLF Proposed to Fix Cyclomatic Complexity's Testing Blind Spots

A software developer has proposed two new code-quality metrics — Coverage Difficulty (CD) and Responsibility Load Factor (RLF) — to address limitations in the widely used Cyclomatic Complexity (CC) measurement. CC, a heuristic dating to the 1970s, treats nested and parallel branches identically, which can underestimate the actual effort required to write exhaustive tests. CD estimates the true number of test cases needed by multiplying nested branch weights and summing parallel ones, rather than simply counting branch points. RLF is calculated by dividing CD by CC, with values significantly above 1.8 signaling that a function is far harder to test than its CC score suggests. The author also recommends refactoring high-RLF functions into flat lookup tables, a pattern they call Cross-Mapped Programming, to reduce hidden testing complexity.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Solana's Account Model Enables Parallel Transactions Unlike Ethereum

Solana processes over 1,000 transactions per second at minimal cost by separating execution logic from data, unlike Ethereum where smart contracts store both together. On Ethereum, this bundled design causes state contention, forcing transactions that touch the same contract to be processed sequentially rather than simultaneously. Solana addresses this through its account model, where executable Programs handle logic while separate non-executable accounts store data. Because each Solana transaction must declare upfront which data accounts it will read or write, the runtime can identify non-overlapping transactions and process them in parallel. This architectural difference is a key reason Solana has gained traction in DeFi and real-world asset tokenization.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Finite State Machines and Zero-Trust Auth Can Make AI Agents More Reliable

Autonomous AI agents often fail in production due to unconstrained behavior, including infinite loops, unauthorized actions, and inconsistent outputs, according to a technical analysis published on tamiz.pro. The piece argues that Large Language Models alone lack built-in mechanisms for strict process control, state management, or secure interaction with external systems. Finite State Machines (FSMs) are proposed as a solution, enforcing deterministic, auditable workflows by restricting agents to well-defined states and transitions at any given moment. Alongside FSMs, the Zero-Trust security model is recommended to govern how agents access resources, applying principles such as least-privilege access and continuous verification. Together, these two paradigms are presented as essential engineering guardrails for building predictable and secure agentic AI systems.

0 comments Read more at DEV Community