Skip to content

TeXada Desktop v0.4.1

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 11 Sep 22:59
v0.4.1
0372ba0

TeXada Desktop 0.4.1

English

  • Change the default local text/planner model to MiniCPM5-2B Q4_K_M. Existing
    saved model selections remain unchanged until explicitly updated.
  • Disable thinking for local Ollama MiniCPM5-2B text requests, including
    planning, generation, completion and retries. Send reasoning_effort: none
    through extra_body for the declared OpenAI SDK range; retain other models'
    and OCR's existing behavior.
  • Match the complete Chinese division phrase “除以” and require a structured
    fraction for explicit Chinese fraction requests, excluding negated requests.
  • Strip complete outer math delimiters from planner tool arguments before
    validation, preserving escaped dollars, internal delimiters and multiple
    math segments.
  • Update the HTTPX2 development dependency and its locked HTTPCore2 transport
    to 2.12.0, addressing the dependency audit findings from the first CI run.
  • Align the example timeout settings with existing 90-second inference and
    240-second request defaults; update current model/deployment documentation
    while keeping historical 1B measurements labeled as such.
  • Record A18 Pro / 8GB local validation:
    six development cases averaged 7.71 seconds, with 6/6 compile/render commits
    and 5/6 strict structure matches; manual mathematical review passed all six.
    This small development set is not a general accuracy benchmark.
  • Validate with 384 passing Python tests and 8 skipped tests, full Ruff and
    diff checks, plus a generation/KaTeX check against a locally rebuilt desktop
    backend. Synchronize Python, JavaScript, Rust, desktop metadata and the UI
    version display to 0.4.1.

Upgrade to v0.4.1, then pull the text model from the official
Hugging Face repository:

ollama pull hf.co/openbmb/MiniCPM5-2B-GGUF:Q4_K_M

If existing settings still select 1B, change Settings → Backend → Text model
to the 2B reference and save. OCR continues to use its separately configured
vision model; text conversion works without installing the OCR model.

中文

  • 默认本地文本与规划模型改为 MiniCPM5-2B Q4_K_M；已保存的模型选择需手动切换。
  • 仅对本地 Ollama 的 MiniCPM5-2B 规划、生成、补全及重试关闭思考，通过 extra_body
    发送 reasoning_effort: none，兼容已声明的 SDK 版本范围；其他模型与 OCR 行为保留。
  • 修复“除以”的不完整替换，并对中文明确分式请求增加结构约束，排除否定指令。
  • 清理工具参数完整的外层数学定界符，保留转义美元、内部定界符和多段内容，避免正确公式
    因外层 $…$ 无法渲染。
  • 将开发依赖 HTTPX2 及其锁定的 HTTPCore2 传输层升级到 2.12.0，修复首轮 CI
    依赖安全审计发现的问题。
  • 示例超时配置同步为现有的推理 90 秒、请求 240 秒；更新当前模型与部署文档，保留历史
    1B 测量的版本说明。
  • 补充 A18 Pro / 8GB 本机实测：六例开发样例
    平均 7.71 秒，6/6 编译渲染并提交，严格结构匹配 5/6，人工数学核对 6/6。
    小样本结果不代表通用准确率。
  • 384 项 Python 测试通过、8 项跳过；相关 Ruff、diff 检查及本机重建桌面后端的真实生成、
    KaTeX 验证通过；Python、JavaScript、Rust、桌面元数据与界面版本统一为 0.4.1。

升级到 v0.4.1 后，从官方 Hugging Face 仓库
拉取文本模型：

ollama pull hf.co/openbmb/MiniCPM5-2B-GGUF:Q4_K_M

若已保存的设置仍选择 1B，请在设置 → 后端 → 文本模型中改为上述 2B 地址并保存。
OCR 继续使用独立配置的视觉模型；不安装 OCR 模型也可使用文本转换。

Installers: signed and notarized macOS arm64/x64 DMGs and Windows x64 NSIS.
Author: CacinieP

Assets 5
Loading