Changelog

What shipped

Every merged pull request since 2026-10-01 that changed something you can run or read: the CLI, the MCP server, the admission hooks, the API, the self-hosted container and the site. Newest first, titles as merged unless a note under one says why it was reworded.

Read from GitHub, npm and PyPI on 2026-10-19: 93 pull requests. A version is linked only if that exact version is published; the main repository is private, so its pull requests are cited by number.

Published now

Merged

2026-10-19

• MCPVendor plan bundle through neurarch-ai/neurarch #355 neurarch-mcp#44
• APISelf-hostedSitecheck: top-level verdict is the worst of the design and every serve/layout/data/spec block #355
• SiteFix routes after #352 + #354: the five new docs 404'd, links to /trust #356
• SiteDeveloper API: document check data/layout/serve, publish OpenAPI 3.1; cross-link /m/ and /changes-feeds/, one nav across generated pages #353
• SiteAppSite: /trust replaces /security, plus /pricing, /changelog and five pre-check docs #354
• SiteHomepage: 0.3.0 evidence on the four answers, and a section for files other automation reads #349
• CLISiteSite hygiene: new social card, llms.txt, public 404 links, www canonical, real 404s #352
• MCP0.19.0: vendor the engine through neurarch-ai/neurarch #348 neurarch-mcp#43neurarch-mcp 0.19.0 on npm
• CLIserve-check: read multimodal configs as the language model, cross-check weights against the checkpoint, add cards and memory flags #348
• CLIneurarch pin-check: Hugging Face downloads not pinned to a commit (aligned with Bandit B615) #346
• CLISiteCLI: pre-check record as a signed in-toto attestation and a CycloneDX ML-BOM fragment #345

2026-10-18

• SiteModel change feeds: Renovate datasource + preset, OSV advisories, shields.io badges #341
• SiteModel change tracker: format-only re-saves, chat template render diff, licence and gating lane #347
• CLIserve-check: vLLM's own CUDA graph estimate structure, the vision tower's profiling pass, and --receipt for what vLLM measured #344
• CLILayout check: non-GEMM time per op, eager vs fused (opt-in, does not clear the gate) #343
• CLILayout check: GEMM shape model (opt-in, does not clear the gate), nccl-tests collectives, three outside throughput tables #342
• MCPDrop a withdrawn layout-accuracy figure before 0.19.0 is published neurarch-mcp#42neurarch-mcp 0.19.0 on npmThe figure compared the layout check with vTrain's measured runs on graphs built with the wrong MLP width. Withdrawn; corrected in #342.
• Hooksneurarch-slurm: ship cli_filter.lua in the wheel, install with --print-lua #340
• MCP0.19.0: per-card layout timing with a simulated 1F1B schedule, engine-parity serving memory neurarch-mcp#41neurarch-mcp 0.19.0 on npm
• CLIHooksRelease prep: neurarch CLI 0.3.0, neurarch-skypilot 0.1.1, neurarch-slurm needs >=0.1.1 #339neurarch 0.3.0 on npm · neurarch-skypilot 0.1.1 on PyPI
• SiteProduct tour re-recorded around the four answers of the change pre-check #338
• CLIHooksSiteChecked-in policy for the pre-checks: .neurarch.yml with presets, owned and expiring waivers, baselines, SARIF #336
• CLISiteLayout check: per-card GEMM efficiency from NeuSight, Primus schedule simulator, vTrain held-out #335The vTrain held-out comparison added here was withdrawn in #342: the graphs used the wrong MLP width.
• CLIserve-check: size memory, quant support and recipes with each engine's own arithmetic, validated against public vLLM logs #337
• CLISiteRelease labels: neurarch CLI 0.2.0 and neurarch-mcp 0.18.0 are on npm #334neurarch 0.2.0 on npm · neurarch-mcp 0.18.0 on npm
• SiteAppHomepage: where the pre-check sits, measured layout evidence under proof, older tour collapsed #332
• AppPricing modal: rename the per-seat app Team plan to App Team #330
• MCP0.18.0: check_parallel_layout and check_serving_config, and the hook pre-checks vLLM / SGLang launches neurarch-mcp#40neurarch-mcp 0.18.0 on npm
• SiteHomepage: See it in action plays the real captures #328
• CLISiteAppPricing, nav and secondary pages: the pre-check for every open-weight model change #327
• Hooksneurarch-skypilot is on PyPI: install with pip #326
• CLIAPILayout check: 1F1B step time, bandwidth-aware head and loss, fp32 logits in memory (stacked on #323) #325
• SiteDaily open-weight model change tracker and /changes #324
• CLILayout check: pipeline A/B on 2 and 4 L4, plus the tied-head fix that missed #319 #323
• CLIAPISelf-hostedServing config pre-check for vLLM, SGLang and llama.cpp launches #322
• SiteAppHomepage: the pre-check for every open-weight model change #321
• CLIAPISelf-hostedParallel layout check: is this TP/PP/DP split legal, balanced, and worth running #319
• MCPClaude Code plugin: server, launch hook and skill in one install neurarch-mcp#39
• Site/m/ model pages: answer "how much GPU memory to fine-tune <model>" in the title, description and H1 #317
• HooksSLURM admission: know what a GPU job needs before it takes a reservation (cli_filter) #316
• CLICLI inference fits says yes/no; data fit reads template graphs #315
• MCP0.17.3: per-card fits / borderline / does not fit, and a data-fit check neurarch-mcp#38neurarch-mcp 0.17.3: not on npm
• MCPdocs(readme): every string in the rightsize and data-fit GIFs is the CLI's own neurarch-mcp#37
• MCPdocs(readme): four more short recordings (Qwen3-8B from HF, a fragmentation OOM, data vs config, rightsize) neurarch-mcp#36
• CLIAPISelf-hostedSiteAppData fit: catch the job that crashes on its first batch, before it takes the card #314
• MCPREADME: open with "Know what a GPU job needs before it takes a reservation" neurarch-mcp#35
• SiteAppHeadline: know what a GPU job needs before it takes a reservation #313
• CLISiteAppMemory: conservative upper bound and a borderline verdict for the training peak #312
• MCPhook: read the .neurarch.yml job block neurarch-bot reads (0.17.2) neurarch-mcp#34neurarch-mcp 0.17.2 on npm
• CLISite/m: read the safetensors header for a tied embedding stored twice (63 -> 65 of 81) #311
• MCPdocs(readme): a See it section with four tool recordings, replacing demo.gif neurarch-mcp#32
• SiteAppHero: three acts on the right-hand card (asked vs needed, an OOM read, a PR caught) #310
• MCPrelease: 0.17.1, six-term memory refit, check_edit, and a Claude Code PreToolUse hook neurarch-mcp#33neurarch-mcp 0.17.1: not on npm
• HooksSkyPilot admission policy: size the model at submit, refuse what cannot fit, cap the cost #309
• SiteLanding: re-aim below the hero at memory, for the platform lead who owns the GPU bill #308
• CLISiteAppMemory: refit the training peak as six physical terms, OOM runs as constraints, checked on 25 new runs #306
• CLIStop two false signals on the demo PR: learned position tables and the App's standing-down comment #305
• AppHero: bot link to the memory-first demo PR (model-ci-example#11) #304
• SiteAppHero demos: size a real model's job; terminal replay on /rightsize #303
• MCPREADME and tool text: current calibration numbers, diagnose_oom in the opening neurarch-mcp#31
• CLISiteModel pages: largest batch per card for full fine-tune, LoRA and serving, with sequence lengths stated #302
• CLISiteAppfeat(demo): /rightsize.html, two commands with real output, replaces the hero's 'Watch a real run' #301
• MCPfeat(oom): diagnose_oom, the tool to call after a CUDA OOM and before changing batch size neurarch-mcp#30
• AppHero: 'See it in action' row points at evidence for the headline #300
• CLISitefix(hf-loader): build each decoder layer's mixer from config; hybrid families match their checkpoints #299
• CLIAPISelf-hostedSiteEdit checks: architecture edits that load, run, and are wrong #298
• CLIfeat(plan-bundle): carry the OOM diagnosis in the plan bundle for neurarch-mcp diagnose_oom #297
• CLISiteAppMemory calibration: 52 L4 runs at the edge of the card, did-not-fit checks 10 to 90 #296
• CLIAppnpx neurarch oom: diagnose a CUDA OOM against the plan, least invasive fix first #293
• CLIAppSizing coverage from the OOM replay: no-index repos, one LoRA peak, --inference #294

2026-10-17

• MCPRelease 0.17.0: refreshed vendored bundles, size_training_job pins eager plan neurarch-mcp#29neurarch-mcp 0.17.0 on npm
• CLIrightsize report: no savings percentage, no card-agreement figure; CLI package text ready for npm #292Title reworded so it does not repeat the withdrawn figure.
• SiteAppPublic memory calibration table: plan vs accelerate estimate-memory vs measured #291
• CLIAppHero card: requested vs needed; rightsize stops pooling memory across small cards #290
• SiteSite: bot, try and README on the reservation line; report priced at $3,000 #289
• SiteSite: two ways to start for clusters #288Title shortened: the rest described fundraising material, which this page does not list.
• SiteAppSite: root hero on the reservation line, per-GPU pricing #287Title shortened: the rest described fundraising material, which this page does not list.
• CLIAPISelf-hostedSiteAppPlan sizes token models at 2048 tokens by default; ConvNeXt, DINOv2, Swin and BEiT readers #286
• MCPplan: accept seq_len and forward it to /api/v1/plan neurarch-mcp#28
• CLIAPISelf-hostedSiteAppAttention default is SDPA in every plan; ViT reads as 197 tokens #285
• MCPfeat(sizing): size_training_job, what a job needs before it takes a GPU neurarch-mcp#27
• CLIfeat(cli): npx neurarch rightsize, requested vs needed GPU over past jobs #283
• CLISitefeat(plan): memory refit, held-out error and card verdict #281Title reworded: the figures it quoted were superseded by the refit in #306 and the published table (#291, #296, #312). Current figures: /memory-calibration.
• CLISitefeat(plan): memory calibrated across the templates, three terms #280Title reworded: the error it quoted was superseded by later refits. Current figures: /memory-calibration.
• CLIAPIfeat(plan): max batch, two-card split, LoRA targets, ComfyUI latent check #279
• CLISitefeat(plan): the fit line counts activations, calibrated on 28 measured runs #276Superseded by later refits. Current figures: /memory-calibration.

2026-10-16

• CLISitefeat(site): name the buyer in the hero, link who-wrote-it, say what the fit line counts #274
• MCPDockerfile builds from source, so a registry building from the repo can start the server neurarch-mcp#26
• MCPfix(trace): probe mode parks saved activations on the CPU neurarch-mcp#25
• MCPfix(trace): mark the edge after a shape-less merge node; drop unconsumed inputs neurarch-mcp#24
• MCPfix(trace): autopatch no longer retains every activation; NEURARCH_TRACE_EXIT for probe processes neurarch-mcp#23
• CLIfix(preflight): hold back flow-dependent shape findings downstream of inferred edges, as lint already does #265
• Sitefeat(datasets): publish 'who wrote the design', refusal rate by author #262