AI SQL Tuner Studio 3.3: Microsoft Store, Native ARM64 and a Query Tuner That Reads the Whole Plan

Release 3.0.5 made sure the AI worked from the right evidence. The releases since then do two things: they make AI SQL Tuner Studio easier to get and run, and they give Query Tuner much more of what the execution plan actually says.

Studio is now in the Microsoft Store. It runs natively on ARM64 PCs. And Query Tuner now reads the plan’s last actual run, the statistics the plan used and the SET options it was compiled with. This post covers releases 3.0.6 through 3.3.2. Here’s what changed.

See it in action: Sample reports

AI SQL Tuner Studio is in the Microsoft Store

Starting with 3.1.0, you can install Studio from the Microsoft Store on Windows 10 and 11. It is still available as a direct download from aisqltuner.com.

A Store install takes care of the parts that used to need attention:

• Updates come through the Store. Help → Check for Updates asks the Store, and Update installs the new version through it.
• No license key. Studio activates from the Microsoft Store account signed in on the PC, including the Free edition. If you also enter a license key, the key takes priority.
• Subscribe from inside the app. In the Store version, Manage license lists Pro, Team and Enterprise with the price for your market and a Subscribe button. Studio activates the subscription as soon as the purchase completes. You manage or cancel it in your Microsoft account.

If activation reports that no Store purchase was found right after you install or buy, the Store may not have reported it yet. Wait a few minutes and use Help → Activate with Microsoft Store. Studio now says this in the message itself.

Store or direct download?

	Microsoft Store	Direct download from aisqltuner.com
Windows 10 and 11 (x64 and ARM64)	Yes	Yes
Windows Server	No (Windows Server has no Store)	Yes
Prerequisites (Windows App Runtime 1.8 and VCLibs)	Already on Windows 11	Already on Windows 11. On Windows Server, Install-AISQLTuner.ps1 installs them
Updates	Through the Store	Help → Check for Updates downloads the new installer
Licensing	Your Microsoft Store account	A license key
30-day money-back guarantee	No	Yes

Install from one place only. Windows treats the two as separate apps, so installing one does not update or replace the other. If you switch, uninstall the one you no longer use.

Native ARM64 support

Studio now runs natively on Windows on ARM PCs. The installer holds both an x64 and an ARM64 package, and Windows installs the one that matches your PC. Before 3.1.0, ARM64 devices ran the x64 app under emulation. We have tested the native build on ARM64 hardware.

This applies to both the Store and the direct download. The direct-download installer is now named AISQLTunerStudio_<version>_x64_ARM64.msixbundle, and Install-AISQLTuner.ps1 installs the Windows App Runtime and VCLibs for the PC’s architecture instead of always for x64.

The install script also works on a clean Windows Server 2022, which we tested. It had been failing there before it reached the app, so if you hit that, this release fixes it. The release now ships the VCLibs packages Studio needs in a Dependencies folder, because Microsoft’s download links serve older versions.

Query Tuner reads the whole plan

In 3.0.5, Query Tuner got its execution plan back on SQL logins. In 3.3, it reads far more of that plan, and of the evidence around it.

What Query Tuner now sees

• The last actual plan. On SQL Server 2019 and later and Azure SQL, when the query has run before and its plan is cached, Query Tuner reads its last actual run from sys.dm_exec_query_plan_stats. Each operator shows estimated against actual rows, and the operator where the estimate first goes wrong is marked. Your query is still never executed.
• Statistics evidence. Each statistic the plan used is listed with its values at compile time, next to its current last update, rows sampled and modifications. Statistics advice now rests on modification counts and sample sizes, not on dates alone.
• Rows read, executions and predicates per operator. A broad seek followed by a residual filter, or a key lookup run thousands of times, is now visible.
• Why a plan is serial. Each statement shows whether its plan is parallel and, if not, the reason the plan gives, such as a scalar function that can’t be inlined. Studio collects server and database-scoped MAXDOP, the cost threshold for parallelism and each function’s inlineability. The AI is told not to recommend more vCores or a higher MAXDOP while such a restriction applies.
• Implicit conversions name the side converted. Each one says whether the column or the parameter is converted, and where. Studio collects each compared column’s data type, length and collation, so the report can name the matching parameter type.
• Parameters and SET options. Each statement lists its compiled parameter values, its local variables and its SET options. Query Store history is split by SET-options context and matched by query hash. The report warns when the application’s options differ from Studio’s, for example ARITHABORT, since Studio’s plan may not be the application’s.
• Batches and procedures split by statement. Each statement’s cost, memory grant and operators are listed separately, the most expensive first.

What was wrong before

• Three plan warnings were never reported. The Cartesian-product (No Join Predicate), filtered-index (Unmatched Indexes) and memory-grant Wait warnings were looked up under names the plan doesn’t use. They are now read as the showplan schema defines them. Stale-statistics columns, exchange spills, spatial guesses and full updates for online index builds are detected for the first time. The same fix applies to the plan summaries in Locking and Blocking Analysis.
• Table row counts were inflated. A table’s rows were counted once per index and again per storage type, so a 200,000-row table could show as 640,000 rows. Each table’s rows are now counted once.
• Estimates are labelled as estimates. The plan summary now says the estimated plan has no operator actuals, and the AI is told not to report estimate errors as measured.

What the benchmark shows

We ran the new Query Tuner through our benchmark before release: 25 runs with GPT-5.6 Luna and GPT-6 Sol at Low, Medium and High reasoning, compared with 20 runs of the previous version.

	Before 3.3	3.3
Runs that returned wrong results	0 of 20	0 of 25
Reports that used the last actual plan or actual rows	4 of 20	25 of 25
Reports that addressed SET options or ARITHABORT	2 of 20	25 of 25
Runs with an UPDATE STATISTICS script	12 of 20	0 of 25

The benchmark database’s statistics have had no modifications since they were built, so leaving them alone is the correct advice.

Two results we want to be clear about. First, workload benefit didn’t change: on this benchmark query, the median cut in total workload time was already 97 to 100 percent for every model and reasoning level, before and after. The query is too easy to separate a better analysis from a good one, so the gains show up in the evidence the reports cite and the scripts they write. Second, the bigger input makes answers slower: about 10 percent slower with GPT-6 Sol and 30 to 45 percent slower with GPT-5.6 Luna at Medium.

The benchmark also caught a defect, which 3.3.2 fixes. Two of five runs recommended ALTER DATABASE SCOPED CONFIGURATION SET AUTO_CREATE_STATISTICS = ON, which SQL Server rejects. Automatic statistics are database options, set with ALTER DATABASE … SET. The analysis input now says how each kind of setting is changed, and the error didn’t recur in 20 further runs. 3.3.2 also stops suggesting you turn on LAST_QUERY_PLAN_STATS when the last actual plan came back without it.

Reports in your language

A new REPORT LANGUAGE dropdown, next to REASONING, lists English and about 30 other languages. You can also type any other language by name or tag, such as Catalan or de-CH. The report’s prose and headings are written in that language. T-SQL scripts, object names, DMV names, wait types and SQL Server messages are not translated, so the scripts still run. Arabic, Hebrew and other right-to-left languages are laid out right to left. English remains the default.

With each analysis sent to the AI SQL Tuner service, Studio now sends the report language and the Windows display language. The service keeps monthly counts per language, not linked to you or your license, to decide which languages to support. Analyses sent to your own AI endpoint send neither.

Amazon RDS for SQL Server (Preview)

Support for Amazon RDS for SQL Server, on Standard, Enterprise and Developer editions, is in preview. Preview means the code is in place and we expect it to work, but we haven’t yet tested it against an actual RDS instance. We expect to do that in early November. If you run Studio against RDS before then, we’d like to hear how it went.

Before this release, RDS was treated as on-premises SQL Server, so reports recommended fixes that need sysadmin or operating system access. Studio now recognizes RDS by its rdsadmin database, and every goal is told it’s on RDS so it gives remedies the RDS way:

• settings and trace flags through the DB parameter group
• cumulative updates by upgrading the RDS engine version
• encrypted connections with rds.force_ssl
• backups as RDS automated backups, and high availability as Multi-AZ

Fix Deadlocks reads deadlocks on RDS from the system_health event files and ring buffer. When a connection fails on the AWS certificate, the message tells you to install the AWS RDS certificate bundle or turn on Trust Server Certificate. Express and Web editions on RDS are not yet supported.

Other improvements

Code Review

• A new Most expensive order reviews the stored procedures, functions and triggers that used the most CPU in Query Store over the last 30 days. Without Query Store it ranks from the plan cache, and failing that, by newest.
• Each definition is sent whole up to 60,000 characters, up from 40,000.

Fix Deadlocks

• The report says what ran each piece of code: host, login and application. For a SQL Agent T-SQL job step, it names the job and step, and says whether the job step or the application code has to change.
• Each module or batch is sent up to 60,000 characters, and up to 300,000 in total.

Server Health Check

• The latest-CU table now comes from the AI SQL Tuner service each time Studio checks your license, so a new CU no longer needs a new release.
• The last CHECKDB date is right under every login language. Under languages that write the day first, such as German or British English, day and month were swapped for the 1st to the 12th. A database that has never had a CHECKDB is now reported as Never Run.

Bring your own AI endpoint

• Primary model now names the deployment Studio calls. A name starting with claude- is called through the Anthropic Messages API, any other name through the Azure OpenAI Responses API.

Security and licensing

• Saved SQL Server passwords and saved BYO AI keys are now encrypted under your Windows account. Earlier versions stored them in plain text.
• A license key activates as the edition it was bought for. When activation is refused, the license panel shows Activation failed and the reason, instead of your old edition.
• Numbers sent for analysis no longer depend on your Windows region format, so 1,234 ms is no longer read as just over one millisecond.

Conclusion

Studio is now in the Microsoft Store and runs natively on ARM64. Query Tuner reads the last actual plan, the statistics behind it and the SET options it was compiled with, and its reports cite that evidence. Reports can be written in your language, and Amazon RDS support is in preview. All editions, including Free, still work with every version of SQL Server, Azure SQL Database, Azure SQL Managed Instance and Fabric SQL Database.

To get AI SQL Tuner Studio 3.3, install it from the Microsoft Store, download the Free edition, or click Help → Check for Updates inside Studio. To see the output first, look at the sample reports. For setup details, see the user guide.