PXA Network
Changelog

Releases

PXA 4.0 is in pre-release: the highlights below come from the ledger-traced changelog (full record in the repo). Downloads are on GitHub releases.

PXA 4.0 highlights in progress

next · 2026-10-15 pre-release
  • +42% over a popular community fork on a 27B, same P100 pair: 26.67 vs 18.78 tok/s tg128.
  • 35B-A3B MoE with MTP on two P100: 138.1 tok/s greedy, +4.5% over the fork in the same bracket and +38.7% over our 10-14 stack.
  • 16k-prompt prefill on the MoE: 1803 tok/s, +23% over the fork (NOT-QUIET, fork in the earlier bracket; branch m2/fa16k).
  • 28k prompts on one V100 (Flash-Next): 324 → 597 tok/s, 1.84x.
  • Tensor split on a Pascal pair: +48% on the 27B (38.0 vs 25.7 tok/s, same output), picked automatically.
  • One V100 on Flash-Next Overdrive: +19.5% prose / +31.5% code (38.8 / 47.5 vs 32.4 / 36.1 tok/s; quiet rerun owed).
  • PXA Control now ships a Windows exe, a macOS Intel build and an Android client, with an agentic Assistant chat covered by a 152-case test suite.

PXA 4.0 · Models in progress

next · 2026-10-15
  • Qwen3.6-35B-A3B (qwen35moe) with its MTP block is supported end to end: encode, boot, 138 tok/s on two P100.
  • Flash-Next fits more hardware.
  • Nemotron, Nemotron-H and Nemotron-H MoE load through a hand port plus a Mamba-2 selective-scan op (CPU + CUDA). Branch m2/nemotron.
  • GLM-5.3-Flash has an N2 file; no 4.0 speed number yet.
  • Ornith 1.5 35B-A3B uses the same MoE paths as the Qwen3.6 MoE.

PXA 4.0 · PXA Control in progress

next · 2026-10-15
  • One download runs everything: a compiled single file (18.8 MB, 0.5 s start) and a Linux `.run` that carries the engine.
  • The UI is rebuilt: Home for first run / ready / serving, grouped navigation, Plugins and Store pages, tier badges.
  • The Assistant is an agent: tools, sub-agents (`spawn_agent`), memory across chats with incognito, recall, `@` references, auto-compaction on exact token counts.
  • A 152-case suite over Control's HTTP API tests the agent loop, found three fixes, plus 28 unit tests. Branch m2/chat-harness.
  • Chat renders math (KaTeX), switches between edited message versions, and fits a phone.
  • The licensed library updates by itself as its own signed, rollback-able artifact, shown beside the encoder's.
  • History is kept on disk and Home Assistant MQTT is opt-in.
  • Hardening H1-H10 ships (loopback session protection, CSRF and rebinding guards, request limits, content-security policy).
  • One-click model downloads, the live header names every server, and the log shows its backlog.

PXA 4.0 · Platforms in progress

next · 2026-10-15
  • Windows x64 Control exe and engine build: six failing CUDA files fixed for MSVC, the engine is `llama-server.exe`, the closed library builds as `ggml-pxqn.dll`. Real-GPU run owed. Branch m2/control-ports.
  • macOS Intel Control build script, with docs that state only what was verified. Branch m2/control-ports.
  • Android client APK (a WebView wrapper that remotes to your Control host; reproducible command-line build; API 34 emulator smoke test). Branch m2/control-ports.
  • Installable web app (PWA) with manifest, shell service worker, host picker, icons.
  • Container fixes: the cache volume is owned by the `pxq` user and the server runs as pid 1.

PXA 4.0 · Licensing / store / sidecar in progress

next · 2026-10-15
  • The PXA Store runs in licd: credits, vouchers, signed catalogue, machine-bound grants, grant-bound downloads. Live since 10-14.
  • A Ko-fi payment becomes a voucher end to end in the local test (store 21/21, base licd 100/100 tests).
  • Control talks to the real store: catalogue, link, voucher, redeem, refresh, unlink, supporter-model downloads.
  • The vLLM sidecar ships in the release tarball and runs without Docker.
  • The sidecar serves PXQN4/PXQN5/PXQN3: V100 TP2 PXQN4 48.0 tok/s; P100 TP2 +8.6% over vLLM PXQ4; PXQN5 top-1 99.84% vs the engine.
  • `pxa-update` fails closed, needs the exact tarball name, stages in the install directory and rolls back.
  • The release package carries the library updater and its verifier, the expert-map tool, the Grafana dashboard and the Overdrive preset.

PXA v3.1

v3.1 · 2026-10-10
  • Flash-Next Overdrive preset on one P100: +14% prose / +18% code.
  • Learned expert map per session; the map against no map: +16% / +25%.
  • pxa-update: apply, rollback; PXA Control installs a newer release.

PXA v3

v3 · 2026-10-05
  • 27B code on one V100 36.7 to 85.8 t/s (PXQN2, speculation as shipped).
  • Flash-Next 32 GB on one P100 plus RAM: 6.5 to 20.4 t/s.
  • PXA 4.0 optimizations

Tensor-split defaults

v2026.09.20 · 2026-09-21
  • Tensor split becomes the default on qualifying identical pairs.

Release candidate 3

v2026.09.13-rc3 · 2026-09-14
  • The automatic decode default is a long n-gram stage alone, on every card family from Pascal up.

Release candidate

v2026.09.07-rc1 · 2026-09-07
  • Public release candidate of the engine line.

v2026.09.02

v2026.09.02 · 2026-09-02
  • PXA 4.0 optimizations
  • PXQ on the CPU, pxq-export, new sm_70 kernels.

v2026.08.31

v2026.08.31 · 2026-08-31
  • The engine folded into the published tree.

PXQ llama.cpp rc3

v2026.08.28-rc3 · 2026-08-28
  • PXQ4 engine release candidate.

rc2

v2026.08.25-rc2 · 2026-08-25
  • PXQ4 engine, two architectures, measured.

v2026.08.21

v2026.08.21 · 2026-08-21
  • Builds and runs correctly without AVX2.