may 26, 2026 to june 25, 2026
showing 100 public activities, 991 private
ba96cdd ci: install web deps with --ignore-scripts + explicit rebuild8e11d02 ci: move pnpm build-script allowlist to pnpm-workspace.yaml163a36f ci: allow esbuild/@swc native build scripts under pnpm 1011405d3 feat(images): set CARGO_TERM_COLOR=always in rust base imagesed5acd9 ci: split build across rust + fedora runners, deploy on fedora-44ea01116 feat(images): install pnpm in ubuntu-24.04 base imageea01116 feat(images): install pnpm in ubuntu-24.04 base imagea7750de ci: install pnpm via npm (gongfoo rust image lacks corepack)3761333 fix: make the workspace pass the CI lint/test gate1b753f9 feat: prerender every route + Gitea Actions deployRefining the "MoE has a smaller win" note above with the actual mechanism, since A3B keeps coming up as a target.
Spec decode's core trick is verification amortization: load the target weights once, push K draft tokens through them in one forward pass, so verifying K tokens ≈ the memory traffic of 1 token. Clean win for dense models (same weights, K positions).
MoE breaks that amortization. The K verification positions can route to different experts, so the step streams the union of experts across all K positions — worst case ≈ K× the expert-weight traffic, i.e. exactly the cost spec decode was trying to amortize away. Shared/attention params still amortize; the expert-MLP traffic largely doesn't.
Net for A3B is a genuine tension, resolvable only by measurement:
Outcome depends on (a) draft acceptance rate and (b) expert-routing dispersion across accepted tokens. Plausible mitigant: adjacent tokens in real sequences often have correlated routing, so the expert union over K positions may be far below worst-case — but that's a "maybe," not a given. The dense 5.1× (Qwen3-8B) does NOT transfer to A3B — do not extrapolate.
The harder blocker is the drafter. EAGLE-3 / dFlash draft heads are trained per target model on its hidden states. So the gating question for A3B is does a matching drafter (EAGLE-3 or dFlash) for Qwen3.6-30B-A3B exist? If no, this is training a draft head for a 30B MoE target (data + compute) — a project, not a spike. Drafter-free variants (n-gram / lookahead / Jacobi) dodge training but have lower acceptance, which compounds the expert-dispersion penalty.
Action: keep the "dense first" ordering. Treat A3B as its own bench cell with its own gate, and make the first check drafter availability — not speedup. No matching drafter → it's a training project, out of scope for the #79 spike.
wednesday, june 24, 2026 — 21:47:10 utcd53e06d hermes: two-stage build, make /opt/hermes writable by uid 100001142929 hermes: single-container deploy (gateway + dashboard), as deployed on bob745a676 hermes: finalize dashboard exposure + local-inference config214850d Add lair/containers image-build repo; hermes as first image7e11a76 Merge feat/F6-beta-polish: public-beta banner + same-origin nginx + deploy (F6)bc7476b feat(F6): public-beta polish — banner, same-origin nginx, deploy notes5600575 Merge feat/B7-upstream-packaging: RPM + systemd + CI for helexa-upstream (B7)452d7d9 feat(B7): packaging + CI for helexa-upstreambc7476b feat(F6): public-beta polish — banner, same-origin nginx, deploy notes5a8f6bc Merge feat/F5-auth-chat: authenticated chat + key usage (F5)508b326 feat(F5): authenticated chat + key usage integration452d7d9 feat(B7): packaging + CI for helexa-upstream21eb211 Merge feat/B6-served-usage: served-usage ledger + reconciliation (B6, #58)f411722 feat(B6): served-usage ledger + reconciliation (#58)508b326 feat(F5): authenticated chat + key usage integration0de99a8 Merge feat/F4-account-dashboard: auth + account dashboard (F4)1bf3348 feat(F4): auth + account dashboard (mockable client)f411722 feat(B6): served-usage ledger + reconciliation (#58)ce29e0c Merge feat/F3-anon-chat: anonymous chat landing + IndexedDB + SSE (F3)c596519 feat(F3): anonymous chat landing — IndexedDB history + SSE + fingerprint1bf3348 feat(F4): auth + account dashboard (mockable client)7c12b9e Merge feat/F2-mission: /mission route — EU digital sovereignty (F2)8dd8277 feat(F2): /mission route — European digital sovereigntyc596519 feat(F3): anonymous chat landing — IndexedDB history + SSE + fingerprinta6b1fdc Merge feat/F1-theming-i18n: theming + 33-lang i18n + usage-ordered selector (F1)7a6f252 feat(F1): theming + 33-language i18n + usage-ordered language selector8dd8277 feat(F2): /mission route — European digital sovereignty8600d4f Merge feat/B5-topup-codes: single-use top-up codes + mint CLI (B5, #59)2348cc2 feat(B5): single-use top-up codes (redeem + mint CLI)7a6f252 feat(F1): theming + 33-language i18n + usage-ordered language selectorbb0d1e5 Merge feat/B3-cortex-upstream-client: cortex upstream entitlement client + chain (B3, #57)a9d7382 feat(B3): cortex upstream entitlement client (#57) + chained provider2348cc2 feat(B5): single-use top-up codes (redeem + mint CLI)f2ba12b Merge feat/B4-account-api: /web/v1 account API + silent fingerprint abuse (B4, #59)d94c62c feat(B4): /web/v1 account API + silent fingerprint multi-account abusea9d7382 feat(B3): cortex upstream entitlement client (#57) + chained providerd94c62c feat(B4): /web/v1 account API + silent fingerprint multi-account abusecb9e7c7 chore: untrack helexa.ai/node_modules + dist (B2 .gitignore slip)2604b9f Merge feat/B2-authz-api: /authz/v1 authority surface + client-auth + sweeper (B2)178e309 Merge feat/F0-helexa-ai-scaffold: helexa.ai frontend scaffold (F0)7907317 feat(F0): helexa.ai frontend scaffold + monorepo coexistencecb9e7c7 chore: untrack helexa.ai/node_modules + dist (B2 .gitignore slip)89ef083 feat(B4): /web/v1 account API + silent fingerprint multi-account abuse2604b9f Merge feat/B2-authz-api: /authz/v1 authority surface + client-auth + sweeper (B2)46befde feat(B2): /authz/v1 authority surface + client-auth + reservation sweeper178e309 Merge feat/F0-helexa-ai-scaffold: helexa.ai frontend scaffold (F0)7907317 feat(F0): helexa.ai frontend scaffold + monorepo coexistence