News
Company news, product releases and engineering practice

PinnedThe SREYUN website is live: five in-house products and reference pricing
The SREYUN website is live, presenting openaiops, netsa, Ruyi OpsBuddy, TokenTally and dlmodel, with the editions and reference prices of each product.

netsa v1.14.2: upgrades and restarts no longer lose logs being processed
v1.14.2 reviews read checkpoints, warm standby, incremental snapshots, AI guardrails and shadow mode, and fixes logs in the processing queue that could be lost on exit.

Ruyi OpsBuddy: a seven-stage diagnosis loop that makes AI actually troubleshoot
OpsBuddy splits troubleshooting into seven stages (detect, hypothesize, gather evidence, conclude, act, verify, learn), with evidence at every step, approval before write operations and automatic verification afterwards.

TokenTally reports: usage first, costs second, with a sign-in-free read-only interactive report
TokenTally's daily, weekly and monthly reports lead with token usage, go to DingTalk, Feishu, WeCom, webhooks and email, and can include a sign-in-free read-only interactive report link.

openaiops Agent OTA: the fleet keeps up after a server upgrade
After a server upgrade there is no need to sign in to each host: online agents that are behind are queued for upgrade automatically, with SHA-256 checks, maintenance windows and automatic rollback.

Tencent Cloud TokenHub bills have no API key: how are key-level costs calculated?
Tencent Cloud bills by inference service and model, without the API key. TokenTally allocates costs to each key exactly as "real bill amount × token share", and the total matches the bill.

Why alerts need evidence review: unified threat level decisions in netsa
A rule hit is not a successful attack. netsa reviews alerts with session evidence, response baselines and account clues, and decides the level in one place, so blocked attacks are not marked critical and real successes are not missed.

Reliable LLM weight downloads on restricted networks: dlmodel's multi-source fallback
dlmodel falls back step by step between sources and within each source: ModelScope ⇁ HF mirror, aria2 multi-connection → resumable HTTP → native SDK, and treats pausing as a first-class state.