Arc 0.51.65 — Measured memory tiers and task improvements

Arc 0.51.65 — Measured VRAM tiers and longer saved work

This Windows x64 development release brings together 0.51.61–0.51.65, following public 0.51.60. It adds optional lower-VRAM tiers, persistent output checkpoints and multi-file website builds. Models stay separate downloads; normal updates preserve models, settings and conversations.

Download page · Feature overview · Detailed workload measurements · Machine-readable evidence · Memory architecture

Choose a memory/speed trade-off

Open Settings → AI Performance → Lower VRAM usage. Tier 1 accepts at most 20% lower measured generation throughput, Tier 2 35%, Tier 3 50%, and Tier 4 75%. The same prompt-throughput limit applies when normal prompt timing is available. These are limits in the short selection test, not a forced delay or a guarantee for every longer task. Tier 4's speed tolerance does not promise 75% less VRAM.

Arc compares supported cache precision and processing-buffer settings on the selected model. Higher tiers can test a small number of CPU-resident blocks when the verified engine supports them. Tier 4 adds at most three directly measured combinations of Q4 cache and CPU placement. Five basic response checks, measured generation speed and memory admission must qualify. Prompt speed is compared when the normal rate is available, and first-token delay when both timings are available; missing timing is not a successful comparison. A higher tier may keep the same setting when more aggressive offload is too slow. Existing successful comparisons are cached. Crashes, out-of-memory results and unhealthy configurations cannot be revived; successful trials rejected only by a stricter speed threshold can be reconsidered at a more tolerant tier.

The model, installed weight quantization and selected context stay unchanged. Turning the option off restores the normal tested profile. No new idle-unload timer is introduced. Strata retains its own engine and is outside these llama.cpp tiers.

Real measurements, with their limits

The linked report compares Normal, Tier 1, Tier 2, Tier 3 and Tier 4 using real inference and workload tools. It distinguishes native token speed from complete task time and whole-device GPU memory from owned worker/server RAM. Source reads, screenshots, vision analysis and file validation count only when their evidence exists. Failures are retained in the report.

The earlier three-tier calibration is retained in the version history with its original scope. The new report separates short calibration from the five-mode workload comparison. These are measurements on one machine, not proof of fit or equivalent performance on 4/8/12/24 GB cards. Lower observed whole-device VRAM during text generation does not establish that a screenshot/vision handoff uses less memory or that every task finishes faster.

Measured on Arc Focus and RTX 5090

One Windows PC, 16,384-token context, three 512-token generation samples per mode. These are raw whole-device text-generation peaks, not exclusive model allocations; pre-load medians were not subtracted.

Mode	Median tok/s	Throughput change	Pre-load GPU median, MiB	Generation GPU peak, MiB	Raw peak MiB saved
Normal	187.28	Reference	1,711	8,278	Reference
Tier 1	185.38	1.01% loss	2,096	7,781	497
Tier 2	184.24	1.62% loss	2,008	7,927	351
Tier 3	182.09	2.77% loss	1,889	7,490	788
Tier 4	52.67	71.87% loss	2,210	7,432	846

All five final modes delivered the requested web answer, readable screenshot/vision result and exact 1,000-line file, and reused the model after work. Manual review still found factual and citation errors. Tier 4's file took 611.56 seconds versus Normal's 121.59 seconds, including recovered model-call failures and an unusually slow call. Its extra raw generation saving over Tier 3 was only 58 MiB in this run, so Extreme was a poor speed/memory trade-off here. Overall workload peaks, vision handoffs, CPU/RAM/disk counters, individual samples, earlier failures and measurement limits are in the full report.

Corrections found during workload testing

• A layout-only screenshot target returns a concrete instruction to choose visible content or request a page viewport. Current-page references beyond the first 100 elements are supported, with the same strict format and live-page checks.
• A fully ended tool response with correctable argument-schema errors keeps the healthy text worker ready for the correction. Malformed or unfinished streams still release the failed worker.
• Newly authored, line-counted JavaScript checks the actual saved line count. A missing section separator is repaired only at a safe, complete-statement boundary; exact-copy requests bypass that repair, and retries do not resave accepted bytes.
• A duplicate-section replay guard checks the current file ID/hash and the same raw section with an unchanged nonempty checkpoint, returning the existing receipt/tail without appending bytes or advancing chunk count. Intentional repetition with an advancing checkpoint and exact-copy requests is preserved. Its 63 focused tests include the diagnostic's 1,000-lines-but-900-unique failure shape.
• Direct-source HTTP 403/429 observations survive compaction so the agent can choose another publisher or an already authorized normal browser route. This does not bypass access restrictions or guarantee another source is available.
• Automatic source selection skips already denied hosts during the same turn. If the available lead pool is blocked, one bounded alternate-source search uses the original query with denied-site exclusions. Explicit user-required access remains distinct; this does not bypass restrictions or guarantee a readable alternative.
• Exact failed-URL guidance and no-tools synthesis for already fulfilled pure web-research contracts address a repeated-URL diagnostic failure. Focused recovery tests passed. Errors remain in the ledger, and a source-read count alone is not a completed answer.
• Combined research/screenshot requests retain a bounded follow-through phase for requested searches and page reads. Receipts distinguish actual retrieved text from search leads and indexed fallback excerpts; multiple URLs do not prove independent corroboration.
• Screenshot guidance favors readable viewport or element captures for text while honoring explicit full-page requests. Vision is instructed to flag unreadable text, with actual visual accuracy reviewed separately from tool completion.
• Capture coverage distinguishes a top-level document-root viewport from a full-page image. Oversized non-root and iframe regions receive guidance to choose a smaller visible target instead of silently widening the screenshot. Focused checks and a readable live capture passed; combined research completion is reviewed separately.
• Untargeted visual-research screenshots use the actual viewport unless the original user explicitly requested whole-page coverage. Receipts describe that scope and suggest scrolling for other sections. Explicit full-page requests, targeted captures and standalone saving keep their existing behavior; whole-page vision may still have legibility limits.
• Screenshot delivery uses the verified saved image and its file card, instead of a webpage URL or invented filename. Requested vision observations stay associated with that image and its recorded source, preserving unreadability caveats.
• Plan-only cleanup preserves substantive answers, citations, mixed code and quoted/fenced text after a Tier 2 diagnostic exposed overly broad removal. The focused suite passed 63 tests and the actual failed answer replay retained its substantive remainder. Successful tool calls do not by themselves establish answer quality; affected source revisions remain diagnostic.
• Final response instructions close the tool phase and retain combined research requirements. A confidently plan-only answer gets one bounded evidence-only composition retry without tools or repeated captures; persistent failure remains incomplete with actual images and receipts preserved. Real explanations, code, quoted plans and user-requested plans remain, and image evidence has a context-sensitive whole-payload bound. The focused repair suite passed 84 tests; this is completion handling, not factual grading.
• Research compaction preserves bounded, query-relevant exact passages and qualifications with omission notices, instead of a navigation-heavy page prefix. Page text remains untrusted evidence; final interpretation still needs accuracy checks.

These describe the implemented fixes. Live success and remaining failures are recorded separately in the workload report.

Six substantial completion guides

Six workflows guide actual source verification, readable screenshots with vision, combined research and image delivery, checked file chunks, debug/verify/save work, and evidence-based recovery. The full references contain 8,100 words in total (1,310–1,390 each), with complete compact task/check/stop recipes. A serialized single-guide selection is 1,811–1,888 characters including the shared notice, identity and formatting; the actual recipe body is smaller. The existing context-dependent budget limits what is loaded; full references are read on demand.

Find them in Settings → Plugins using the guide index. They use existing tools and permissions, preserve saved off switches, and do not add new executable backends. Routing distinguishes saved browser screenshots from actual live sharing. The final matrix records selected guide IDs and guidance hashes; document counts are not proof of intelligence, performance or output accuracy.

Longer output, saved as it goes

Ultra's optional Until done mode writes large files in revision-checked sections, using saved requirements, the current file revision, a structure outline and an exact ending to continue. Smaller contexts receive smaller sections and output reserves. Incremental JSON validation, checked record batches, explicit count checks and repetition protection provide bounded checks; they do not establish that every requested line, behavior or fact is correct. Bounded retries preserve accepted work, and the workload report records where continuation still failed.

Persistent task checkpoints and pinned requirements can survive a restart. A resume checks the current file hash before continuing. This uses storage to preserve work; the model still has a finite context window and may need multiple generations.

Build a website one file at a time

Multi-file website projects save a shared plan, dependency contracts, pages, styles, scripts and nested directories in the local coding workspace. Arc works on one file at a time instead of replaying every completed source file in the next prompt. It checks supported syntax and local links, attempts bounded repairs and exports the directory with file-hash verification.

Open Code → Local website workspace → Export & open folder for the finished directory. Source creation does not automatically execute or publish generated code. Existing File History, single-file editing, protected PHP preview and reviewed SSH publishing remain available.

RAM, SSD and GPU are different resources

The default llama.cpp text-inference RAM ceiling remains 32 GiB, reduced further by physical/free-memory reserves. It does not cap the entire PC, every browser tab or separate voice/vision processes. Low-VRAM mode can reduce supported caches/buffers or move eligible computation to CPU/RAM. It does not turn the SSD into directly usable VRAM.

Strata's separate SSD-aware expert engine is unchanged. Saved file checkpoints and retrieval also use local storage, but are not SSD expert inference or unlimited active context. The architecture guide explains these boundaries and model residency during vision handoffs.

Verification and downloads

Download Arc-Setup-0.51.65.exe from this release's assets. Size: 334,808,652 bytes. SHA-256:

f5397fc3aa9515654d9aac0de14674537921e9d8e805258bc413b5bbd2ac993b

The release includes checksums and the benchmark report. The installer is not Authenticode-signed. No matching VirusTotal scan has been supplied for this exact hash; earlier scans apply only to their own files. Arc's signed update metadata and checksum verification are separate from Windows publisher signing.

Final test and workload coverage is recorded in the linked report. Short response checks, automated fixtures and live tasks have different scopes; none establishes universal model accuracy or reliability. Browser sign-ins, CAPTCHA handling, external-site access and available hardware can still limit a task. The new tiers add no runtime dependency, model download or automatic context reduction.

The latest completed full automated regression suite passed 3,453 of 3,458 tests with 5 skipped and 0 failures, in 556.146 seconds, before the subsequent finalization repair. After that repair, 478 tests across 27 affected files passed with no skips in 11.134 seconds. Overlapping counts are not added. The report preserves an earlier concurrent crawl timeout and its successful isolated rerun. These counts describe automated checks; final live task accuracy and installer identity are verified separately.