Skip to content

XeStrata xe0.1.45-1

Choose a tag to compare

@MistVVK MistVVK released this 18 Oct 12:49
· 3183 commits to main since this release

Warning

This is an old release. Get the latest from the latest release.
古いリリースです。最新版は最新のリリースから入れてください。

English | 日本語

English

Install on Linux

1. Podman

Only Podman and podman-compose go on the host; XeStrata goes into a Fedora 44 container, from this release's rpm (rootless).

git clone https://github.com/MistVVK/XeStrata && cd XeStrata/container
mkdir -p ~/.config/xestrata ~/.local/state/xestrata ~/.local/share/xestrata
podman compose build                         # the image, with the latest release and the GPU libraries
podman compose run --rm xestrata --setup     # answer the questions; it downloads the model
podman compose up -d                         # http://127.0.0.1:8095 (podman compose down stops it)
  • GPU: Intel's as it is. For AMD add -f compose.yaml -f compose.amd.yaml after podman compose, for NVIDIA -f compose.yaml -f compose.nvidia.yaml (NVIDIA needs the NVIDIA Container Toolkit's CDI specification on the host). The container has cuda13.4, so Pascal and Volta are not used.
  • What the host needs: the GPU's driver, and your user in the render group (sudo usermod -aG render $USER).

2. apt / dnf

  1. Pick the package for your GPU from the table below and download it (with SHA256SUMS).
  2. Check it: sha256sum -c SHA256SUMS --ignore-missing.
  3. Install it: sudo apt install ./xestrata-contrib-cuda13.1_0.1.45-1_amd64.deb (Ubuntu) or sudo dnf install ./xestrata-contrib-cuda13.4-0.1.45-1.x86_64.rpm (Fedora).
  4. Run xestrata --setup in a terminal. If a GPU library or the render group is missing, it prints the command that installs it and stops; run that command, then xestrata --setup again.

Note

Without an NVIDIA GPU, either CUDA version works the same.

GPU Ubuntu 26.04 Fedora 44
Recommended: Intel, AMD (RDNA2 or later), NVIDIA (Turing or later) xestrata-contrib-cuda13.1 xestrata-contrib-cuda13.4
Intel, AMD (RDNA2 or later), NVIDIA Pascal and Volta (P100, P40, GTX 10 series, V100 and others) xestrata-contrib-cuda12.4 none

Install on Windows (new)

  1. Download XeStrata-0.1.45-1-windows-x64.7z (with SHA256SUMS) and unpack it into a folder of your choice.
  2. Start xestrata.exe. It takes you through the install (the GPU and the model, the download) and the start on screen. On the console, xestrata-cli.exe (used as xestrata on Linux).
  • What it needs: the .NET 10 runtime, the Visual C++ Redistributable and the GPU's driver.
  • GPUs: Intel and NVIDIA (NVIDIA from Pascal on). AMD GPUs do not run on Windows yet.
  • Where things go: nothing is written into the 7z's folder. The settings go to %APPDATA%\xestrata, the logs and the rest to %LOCALAPPDATA%\xestrata, the models to %LOCALAPPDATA%\xestrata\data.

Also in this release

  • There is no free-software-only package (xestrata-free); build it from the source as docs/BUILD.md describes.
  • The version changed, so conversations saved with --conversation-save by an earlier version are not read.
  • This release carries upstream Strata up to v0.1.45 (xe0.1.44.1-1 was withdrawn before its speed check; its changes are here too).

New

  • Image generation with XeStrata's own engine (Qwen-Image-2.1-Turbo, optional: xestrata --setup --images qwen-image). One 1024×1024 image of 8 steps: 7.7 s on the Arc Pro B70, 13.3 s on the RTX 4070, 22.2 s on the RX 9060 XT. It also changes or reuses up to 3 reference images (POST /v1/images/edits, or attached on the web page), and a chat can make and edit images itself (the web page, and the Responses API's image_generation tool).
  • Windows: the 7z above, with a graphical installer and launcher.
  • Qwen3.6-35B-A3B (BETA): a thinking model with its MTP layer, for PCs with 31 GB of RAM or more; and Unsloth's 5-, 6- and 8-bit files of Qwen3.8-Flash-Next (UD-Q5_K_XL, UD-Q6_K_XL, Q8_0).
  • Two GPUs: --layer-split auto also prices the second GPU as a cache of the experts the first one lacks, and takes the faster of that and the layer split. Arc Pro B70 + RTX 4070, IQ3_XXS 4K: decode 73.8 → 84.7 tok/s.
  • Podman: XeStrata in a container, with only Podman on the host (README, "Podman").

Faster

  • Short prompts on the Arc Pro B70: the prompt's expert products read the experts already in VRAM where they lie, and each layer takes the int8 or the FP16 products, whichever this GPU measured faster for that many rows. IQ3_XXS 1K prefill: 872 → 1,170 tok/s.
  • Decode on the RX 9060 XT: faster than upstream at every length (+1% to +8%).
  • Two GPUs: the second GPU no longer holds 1 GiB out of its expert cache, and its prompt rows are handed over in pieces. Two 10 GB cards: 560-700 more experts on the second one, prefill +4-5%.
  • Conversations: a new conversation that starts like a saved one reuses that one's checkpoint inside the conversation, and the saved one is kept.

Fixed

  • Two runs of the same prompt give the same answer: setup measures the free VRAM and the prompt's CPU share once and writes them, and the CPU and the GPU compute an expert to the same bits.
  • serve: max_tokens counts the thinking too, and Anthropic's thinking.budget_tokens caps it; a request without sampling settings takes the model card's.
  • Windows Subsystem for Linux 2: setup finds the GPUs there.
  • A prompt of fewer than 16 tokens no longer stops Ornith-1.5 and llm-jp-4.1's prompt path on a small GPU.

Speed against upstream

XeStrata xe0.1.45-1 against upstream v0.1.45 on the same GPU, with Qwen3.8-Flash-Next IQ3_XXS. generate reads a prompt of 1K to 262K tokens and writes 256 tokens, both engines following the same continuation; serve is a 32K conversation with strata --serve; everyday use is a Japanese chat (chat-ja) and a coding agent (agent) through serve/server.py; image generation is one 1024×1024 image of 8 steps.

Note

  • generate: temperature 0 (greedy). Within ±1% counts as ≈ equal.
  • serve and everyday use: temperature 0.7, each engine writing its own text, so a run can land about 10% off the median. Within ±1% for prefill and ±5% for the wait, decode and first token counts as ≈ equal.
  • image generation: the time of an image from a fixed prompt and seed, each engine as its own setup sets it up on the GPU. Within ±1% counts as ≈ equal.
  • All are the median of two alternating runs.

Intel Arc Pro B70 (32 GB): upstream v0.1.45 is its sycl/ port

generate (temperature 0, tokens/s)

Type and length prefill XeStrata xe0.1.45-1 prefill upstream v0.1.45 Difference decode XeStrata xe0.1.45-1 decode upstream v0.1.45 Difference
IQ3_XXS 1K 1,171 679 ✅ ×1.7 80.5 57.7 ✅ +40%
IQ3_XXS 4K 1,593 1,262 ✅ ×1.3 74.7 56.5 ✅ +32%
IQ3_XXS 32K 2,345 1,202 ✅ ×2.0 88.5 60.8 ✅ +46%
IQ3_XXS 64K 2,427 1,154 ✅ ×2.1 82.4 60.6 ✅ +36%
IQ3_XXS 128K 2,411 1,044 ✅ ×2.3 69.5 60.9 ✅ +14%
IQ3_XXS 262K 2,299 892 ✅ ×2.6 75.7 54.3 ✅ +39%

serve and everyday use (temperature 0.7)

Measure XeStrata xe0.1.45-1 upstream v0.1.45 Difference
serve 32K first prefill (tok/s) 2,329 1,195 ✅ ×1.9
serve 32K wait for a follow-up (ms) 160 297 ✅ −46%
serve 32K decode (tok/s) 80.7 55.4 ✅ +46%
chat-ja first token (s) 0.42 0.75 ✅ −44%
chat-ja decode (tok/s) 63.1 43.4 ✅ +45%
agent first token (s) 2.70 3.50 ✅ −23%
agent prefill (tok/s) 1,466 1,010 ✅ ×1.5
agent decode (tok/s) 74.3 57.5 ✅ +29%

image generation (1024×1024, 8 steps, seconds)

Measure XeStrata xe0.1.45-1 (strata-image, Q8_0) upstream v0.1.45 (Vulkan, Q8_0) Difference
An image 7.7 51.6 ✅ ×6.7
Sampling 6.9 46.2 ✅ ×6.6
VAE 0.4 3.0 ✅ ×6.9

NVIDIA GeForce RTX 4070 (12 GB): upstream v0.1.45 is its CUDA build

generate (temperature 0, tokens/s)

Type and length prefill XeStrata xe0.1.45-1 prefill upstream v0.1.45 Difference decode XeStrata xe0.1.45-1 decode upstream v0.1.45 Difference
IQ3_XXS 1K 299 258 ✅ ×1.2 56.9 46.4 ✅ +23%
IQ3_XXS 4K 543 526 ✅ ×1.0 52.5 45.7 ✅ +15%
IQ3_XXS 32K 1,167 1,143 ✅ ×1.0 61.7 50.1 ✅ +23%
IQ3_XXS 64K 1,092 1,071 ✅ ×1.0 56.3 46.9 ✅ +20%
IQ3_XXS 128K 1,064 1,044 ✅ ×1.0 54.0 43.6 ✅ +24%
IQ3_XXS 262K 1,018 997 ✅ ×1.0 52.2 43.1 ✅ +21%

serve and everyday use (temperature 0.7)

Measure XeStrata xe0.1.45-1 upstream v0.1.45 Difference
serve 32K first prefill (tok/s) 1,166 1,136 ✅ ×1.0
serve 32K wait for a follow-up (ms) 228 269 ✅ −15%
serve 32K decode (tok/s) 71.3 59.4 ✅ +20%
chat-ja first token (s) 0.76 0.90 ✅ −17%
chat-ja decode (tok/s) 64.3 55.9 ✅ +15%
agent first token (s) 7.24 7.45 ≈ equal (−3%)
agent prefill (tok/s) 548 534 ✅ ×1.0
agent decode (tok/s) 64.0 60.3 ✅ +6%

image generation (1024×1024, 8 steps, seconds)

Measure XeStrata xe0.1.45-1 (strata-image, Q6_K) upstream v0.1.45 (CUDA, Q6_K) Difference
An image 13.3 25.8 ✅ ×1.9
Sampling 11.4 18.8 ✅ ×1.6
VAE 0.7 3.9 ✅ ×5.4

AMD Radeon RX 9060 XT (16 GB): upstream v0.1.45 is its HIP build

generate (temperature 0, tokens/s)

Type and length prefill XeStrata xe0.1.45-1 prefill upstream v0.1.45 Difference decode XeStrata xe0.1.45-1 decode upstream v0.1.45 Difference
IQ3_XXS 1K 649 308 ✅ ×2.1 50.9 47.6 ✅ +7%
IQ3_XXS 4K 1,009 372 ✅ ×2.7 46.3 45.3 ✅ +2%
IQ3_XXS 32K 1,245 395 ✅ ×3.2 52.7 52.1 ✅ +1%
IQ3_XXS 64K 1,236 388 ✅ ×3.2 50.5 48.2 ✅ +5%
IQ3_XXS 128K 1,210 390 ✅ ×3.1 46.5 43.2 ✅ +8%
IQ3_XXS 262K 1,146 383 ✅ ×3.0 44.1 41.2 ✅ +7%

serve and everyday use (temperature 0.7)

Measure XeStrata xe0.1.45-1 upstream v0.1.45 Difference
serve 32K first prefill (tok/s) 1,102 400 ✅ ×2.8
serve 32K wait for a follow-up (ms) 249 323 ✅ −23%
serve 32K decode (tok/s) 53.3 51.4 ≈ equal (+4%)
chat-ja first token (s) 0.71 1.01 ✅ −29%
chat-ja decode (tok/s) 48.6 46.9 ≈ equal (+4%)
agent first token (s) 5.59 10.89 ✅ −49%
agent prefill (tok/s) 706 364 ✅ ×1.9
agent decode (tok/s) 50.7 51.5 ≈ equal (−2%)

image generation (1024×1024, 8 steps, seconds)

Measure XeStrata xe0.1.45-1 (strata-image, Q8_0) upstream v0.1.45 (HIP, Q8_0) Difference
An image 22.2 73.5 ✅ ×3.3
Sampling 20.0 55.3 ✅ ×2.8
VAE 1.5 11.0 ✅ ×7.3

日本語

Linux に入れる

1. Podman

ホストに入れるのは Podman と podman-compose だけ。XeStrata は Fedora 44 のコンテナの中に、このリリースの rpm から入る(rootless)。

git clone https://github.com/MistVVK/XeStrata && cd XeStrata/container
mkdir -p ~/.config/xestrata ~/.local/state/xestrata ~/.local/share/xestrata
podman compose build                         # 最新のリリースと GPU のライブラリを入れたイメージを作る
podman compose run --rm xestrata --setup     # 質問に答えてモデルを取ってくる
podman compose up -d                         # http://127.0.0.1:8095(止めるときは podman compose down)
  • GPU: Intel はこのまま。AMD は -f compose.yaml -f compose.amd.yaml、NVIDIA は -f compose.yaml -f compose.nvidia.yaml を podman compose の後に付ける(NVIDIA はホストに NVIDIA Container Toolkit の CDI の設定が要る)。コンテナの中は cuda13.4 なので、Pascal と Volta は使えない。
  • ホストに要るもの: GPU のドライバーと、自分が render グループに入っていること(sudo usermod -aG render $USER)。

2. apt / dnf

  1. 下の表から GPU に合うパッケージを選び、SHA256SUMS と一緒にダウンロードする。
  2. 確かめる: sha256sum -c SHA256SUMS --ignore-missing
  3. 入れる: sudo apt install ./xestrata-contrib-cuda13.1_0.1.45-1_amd64.deb(Ubuntu)または sudo dnf install ./xestrata-contrib-cuda13.4-0.1.45-1.x86_64.rpm(Fedora)
  4. 端末で xestrata --setup を実行する。GPU のライブラリや render グループが足りなければ、入れるコマンドを表示して止まるので、そのコマンドを実行してから、もう一度 xestrata --setup を実行する。

Note

NVIDIA の GPU がなければ、どちらの CUDA の版でも同じように動きます。

GPU Ubuntu 26.04 Fedora 44
おすすめ: Intel、AMD(RDNA2 以降)、NVIDIA(Turing 以降) xestrata-contrib-cuda13.1 xestrata-contrib-cuda13.4
Intel、AMD(RDNA2 以降)、NVIDIA の Pascal と Volta(P100、P40、GTX 10 シリーズ、V100 など) xestrata-contrib-cuda12.4 なし

Windows に入れる(新しく)

  1. XeStrata-0.1.45-1-windows-x64.7z を SHA256SUMS と一緒にダウンロードし、好きなフォルダーに展開する。
  2. xestrata.exe を起動する。インストール(GPU とモデルの選択、ダウンロード)と起動を画面で案内します。コンソールでは xestrata-cli.exe(Linux の xestrata と同じ使い方)。
  • 必要なもの: .NET 10 のランタイム、Visual C++ 再頒布可能パッケージ、GPU のドライバー。
  • GPU: Intel と NVIDIA(NVIDIA は Pascal 以降)。AMD の GPU は、Windows ではまだ動きません。
  • 置き場所: 7z のフォルダーには何も書きません。設定は %APPDATA%\xestrata、ログなどは %LOCALAPPDATA%\xestrata、モデルは %LOCALAPPDATA%\xestrata\data です。

ほかに

  • このリリースには自由ソフトウェアだけのパッケージ(xestrata-free)はありません。docs/BUILD.ja.md のとおりソースから作れます。
  • 版が変わったので、前の版の --conversation-save で保存した会話は読みません。
  • このリリースは upstream の Strata v0.1.45 までを取り込んでいます(xe0.1.44.1-1 は速さの確認の前に取り下げました。その変更もここに入っています)。

新しくできること

  • XeStrata 自前のエンジンでの画像の生成(Qwen-Image-2.1-Turbo、入れるかは選べます: xestrata --setup --images qwen-image)。1024×1024・8 ステップの 1 枚は、Arc Pro B70 で 7.7 秒、RTX 4070 で 13.3 秒、RX 9060 XT で 22.2 秒です。参照画像(3 枚まで)を変えたり使ったりもできます(POST /v1/images/edits、Web のページでの添付)。チャットの中でも、モデルが自分で画像を作って直せます(Web のページ、Responses API の image_generation の道具)。
  • Windows: 上の 7z。画面で案内するインストーラーと起動の窓口が付きます。
  • Qwen3.6-35B-A3B(BETA): MTP の層付きの思考するモデルで、RAM 31 GB 以上の PC 向けです。Qwen3.8-Flash-Next の Unsloth の 5・6・8 ビットのファイル(UD-Q5_K_XL・UD-Q6_K_XL・Q8_0)も選べます。
  • 2 枚の GPU: --layer-split auto は、2 枚目を層の分担にする場合に加えて、1 枚目が持たないエキスパートの置き場にする場合も見積もり、速い方を選びます。Arc Pro B70 + RTX 4070、IQ3_XXS の 4K: decode 73.8 → 84.7 tok/s。
  • Podman: ホストには Podman だけを入れ、XeStrata をコンテナで動かせます(README の「Podman」)。

速くなったこと

  • Arc Pro B70 での短いプロンプト: プロンプトのエキスパートの積は、VRAM にあるエキスパートをその場所から読みます。層ごとに、int8 と FP16 の積のうち、その行数でこの GPU が測って速い方を使います。IQ3_XXS の 1K の prefill: 872 → 1,170 tok/s。
  • RX 9060 XT の decode: すべての長さで upstream より速くなりました(+1〜+8%)。
  • 2 枚の GPU: 2 枚目がエキスパートのキャッシュから 1 GiB を余分に引かなくなり、プロンプトの行も分けて渡します。10 GB のカード 2 枚で、2 枚目に置けるエキスパートが 560〜700 増え、prefill +4〜5%。
  • 会話: 保存した会話と同じ始まりの新しい会話は、その会話の途中のチェックポイントを使い回し、保存した会話も残します。

直したこと

  • 同じプロンプトを 2 回送ると同じ答えになります。setup が空いている VRAM とプロンプトの CPU の分担を一度測って書き、CPU と GPU がエキスパートを同じビットで計算します。
  • serve: max_tokens に思考も数え、Anthropic の thinking.budget_tokens も思考の上限にします。サンプリングの設定のない要求には、モデルカードの値を使います。
  • WSL2(Windows Subsystem for Linux 2): setup が GPU を見つけます。
  • 小さい GPU で、16 トークン未満のプロンプトのとき Ornith-1.5・llm-jp-4.1 のプロンプトの経路が止まらなくなりました。

upstream との速さの比較

同じ GPU で、XeStrata xe0.1.45-1 と upstream v0.1.45 を Qwen3.8-Flash-Next IQ3_XXS で比べました。generate は 1K〜262K トークンの prompt を読んで 256 トークンを書きます(両方のエンジンが同じ続きの文をなぞる)。serve は strata --serve での 32K の会話、通常使用は serve/server.py を通した日本語のチャット(chat-ja)とコーディングエージェント(agent)、画像の生成は 1024×1024・8 ステップの 1 枚です。

Note

  • generate:温度 0(greedy)。差が ±1% 以内なら ≒ 同等です。
  • serve と通常使用:温度 0.7 で、それぞれのエンジンが自分の文を書くので、1 回ごとに中央値から 10% 前後ずれることがあります。差が prefill は ±1% 以内、待ち・decode・最初のトークンは ±5% 以内なら ≒ 同等です。
  • 画像の生成:決まったプロンプトと種で 1 枚ずつ作った時間です。それぞれのエンジンの setup がその GPU に選ぶ作り方で動かします。差が ±1% 以内なら ≒ 同等です。
  • どれも、交互に 2 回ずつ測った中央値です。

Intel Arc Pro B70 (32 GB): upstream v0.1.45 は sycl/ 版

generate(温度 0、トークン/秒)

形式・長さ prefill XeStrata xe0.1.45-1 prefill upstream v0.1.45 差 decode XeStrata xe0.1.45-1 decode upstream v0.1.45 差
IQ3_XXS 1K 1,171 679 ✅ 1.7 倍 80.5 57.7 ✅ +40%
IQ3_XXS 4K 1,593 1,262 ✅ 1.3 倍 74.7 56.5 ✅ +32%
IQ3_XXS 32K 2,345 1,202 ✅ 2.0 倍 88.5 60.8 ✅ +46%
IQ3_XXS 64K 2,427 1,154 ✅ 2.1 倍 82.4 60.6 ✅ +36%
IQ3_XXS 128K 2,411 1,044 ✅ 2.3 倍 69.5 60.9 ✅ +14%
IQ3_XXS 262K 2,299 892 ✅ 2.6 倍 75.7 54.3 ✅ +39%

serve と通常使用(温度 0.7)

計測 XeStrata xe0.1.45-1 upstream v0.1.45 差
serve 32K 最初の prefill (tok/s) 2,329 1,195 ✅ 1.9 倍
serve 32K 会話の続きの待ち (ms) 160 297 ✅ −46%
serve 32K decode (tok/s) 80.7 55.4 ✅ +46%
chat-ja 最初のトークン (s) 0.42 0.75 ✅ −44%
chat-ja decode (tok/s) 63.1 43.4 ✅ +45%
agent 最初のトークン (s) 2.70 3.50 ✅ −23%
agent prefill (tok/s) 1,466 1,010 ✅ 1.5 倍
agent decode (tok/s) 74.3 57.5 ✅ +29%

画像の生成(1024×1024・8 ステップ、秒)

計測 XeStrata xe0.1.45-1(strata-image、Q8_0) upstream v0.1.45(Vulkan、Q8_0) 差
1 枚 7.7 51.6 ✅ 6.7 倍
サンプリング 6.9 46.2 ✅ 6.6 倍
VAE 0.4 3.0 ✅ 6.9 倍

NVIDIA GeForce RTX 4070 (12 GB): upstream v0.1.45 は CUDA 版

generate(温度 0、トークン/秒)

形式・長さ prefill XeStrata xe0.1.45-1 prefill upstream v0.1.45 差 decode XeStrata xe0.1.45-1 decode upstream v0.1.45 差
IQ3_XXS 1K 299 258 ✅ 1.2 倍 56.9 46.4 ✅ +23%
IQ3_XXS 4K 543 526 ✅ 1.0 倍 52.5 45.7 ✅ +15%
IQ3_XXS 32K 1,167 1,143 ✅ 1.0 倍 61.7 50.1 ✅ +23%
IQ3_XXS 64K 1,092 1,071 ✅ 1.0 倍 56.3 46.9 ✅ +20%
IQ3_XXS 128K 1,064 1,044 ✅ 1.0 倍 54.0 43.6 ✅ +24%
IQ3_XXS 262K 1,018 997 ✅ 1.0 倍 52.2 43.1 ✅ +21%

serve と通常使用(温度 0.7)

計測 XeStrata xe0.1.45-1 upstream v0.1.45 差
serve 32K 最初の prefill (tok/s) 1,166 1,136 ✅ 1.0 倍
serve 32K 会話の続きの待ち (ms) 228 269 ✅ −15%
serve 32K decode (tok/s) 71.3 59.4 ✅ +20%
chat-ja 最初のトークン (s) 0.76 0.90 ✅ −17%
chat-ja decode (tok/s) 64.3 55.9 ✅ +15%
agent 最初のトークン (s) 7.24 7.45 ≒ 同等(−3%)
agent prefill (tok/s) 548 534 ✅ 1.0 倍
agent decode (tok/s) 64.0 60.3 ✅ +6%

画像の生成(1024×1024・8 ステップ、秒)

計測 XeStrata xe0.1.45-1(strata-image、Q6_K) upstream v0.1.45(CUDA、Q6_K) 差
1 枚 13.3 25.8 ✅ 1.9 倍
サンプリング 11.4 18.8 ✅ 1.6 倍
VAE 0.7 3.9 ✅ 5.4 倍

AMD Radeon RX 9060 XT (16 GB): upstream v0.1.45 は HIP 版

generate(温度 0、トークン/秒)

形式・長さ prefill XeStrata xe0.1.45-1 prefill upstream v0.1.45 差 decode XeStrata xe0.1.45-1 decode upstream v0.1.45 差
IQ3_XXS 1K 649 308 ✅ 2.1 倍 50.9 47.6 ✅ +7%
IQ3_XXS 4K 1,009 372 ✅ 2.7 倍 46.3 45.3 ✅ +2%
IQ3_XXS 32K 1,245 395 ✅ 3.2 倍 52.7 52.1 ✅ +1%
IQ3_XXS 64K 1,236 388 ✅ 3.2 倍 50.5 48.2 ✅ +5%
IQ3_XXS 128K 1,210 390 ✅ 3.1 倍 46.5 43.2 ✅ +8%
IQ3_XXS 262K 1,146 383 ✅ 3.0 倍 44.1 41.2 ✅ +7%

serve と通常使用(温度 0.7)

計測 XeStrata xe0.1.45-1 upstream v0.1.45 差
serve 32K 最初の prefill (tok/s) 1,102 400 ✅ 2.8 倍
serve 32K 会話の続きの待ち (ms) 249 323 ✅ −23%
serve 32K decode (tok/s) 53.3 51.4 ≒ 同等(+4%)
chat-ja 最初のトークン (s) 0.71 1.01 ✅ −29%
chat-ja decode (tok/s) 48.6 46.9 ≒ 同等(+4%)
agent 最初のトークン (s) 5.59 10.89 ✅ −49%
agent prefill (tok/s) 706 364 ✅ 1.9 倍
agent decode (tok/s) 50.7 51.5 ≒ 同等(−2%)

画像の生成(1024×1024・8 ステップ、秒)

計測 XeStrata xe0.1.45-1(strata-image、Q8_0) upstream v0.1.45(HIP、Q8_0) 差
1 枚 22.2 73.5 ✅ 3.3 倍
サンプリング 20.0 55.3 ✅ 2.8 倍
VAE 1.5 11.0 ✅ 7.3 倍