Summary A local evaluation finds Qwen’s default extra-high reasoning mode consumes context and latency on trivial prompts.
The details
The supplied evaluation describes Qwen 3.8 27B as an Apache 2.0-licensed, vision-capable 27B-parameter model with 262,144-token maximum context length.
On 17GB Q4_K_M GGUF, an SVG prompt reportedly used 22,276 reasoning tokens, produced 3,223, and took 21 minutes at extra-high reasoning.
With reasoning disabled, the same SVG prompt reportedly took 137 seconds.
The evaluator says LM Studio’s 8,192-token context default was exhausted by mundane-task reasoning; the 262,144-token maximum avoided this failure mode.
The source reports bounding boxes [195, 290, 370, 780] and [445, 320, 675, 850] on a 0-to-1000 scale, without independent benchmark comparison.
Blume makes documentation machine-readable by default
Summary Blume turns Markdown docs into static sites with llms.txt, Markdown negotiation, and optional MCP access.
The details
Blume v1.5.1 is a free, open-source Markdown-first documentation framework with built-in search, theming, SEO, internationalization, and Astro or Vite deployment.
By default, it generates llms.txt as a navigable summary index and llms-full.txt as a full Markdown corpus; draft pages are excluded.
Appending .md to a documentation URL returns Markdown, while Accept: text/markdown negotiates it on supported server deployments.
Its optional hosted MCP server exposes four read-only tools: search_docs, get_page, list_pages, and get_navigation.
Summary Rust MCP server turns lifecycle observations into shared Markdown wiki and bounded handoffs.
The details
ai-memory compiles sanitized lifecycle observations into a Git-backed Markdown wiki; FTS5 and rule-based summarization run in zero-LLM mode.
Prompts and post-compaction captures max 16 KiB; notifications and tool excerpts max 2 KB; every observation retains a 16 KiB durable backstop.
`ai-memory run claude` and `ai-memory run codex --yolo` continue one managed workstream with a portable visible-event ledger and native harness resume.
Codex lacks an automatic true session-end hook, so it requires `ai-memory finalize-session` for a final summary; Claude Code supports lifecycle hooks and optional session-aware isolation.
oMLX Adds Tiered Local Inference for Apple Silicon
Summary The macOS server combines continuous batching with persistent SSD-backed KV caching behind an OpenAI-compatible API.
The details
oMLX serves text LLMs, VLMs, OCR models, embedding models, and rerankers through an OpenAI-compatible endpoint at http://localhost:8000/v1.
Its block-based hot-memory and cold-SSD KV cache supports prefix sharing and restores matching cached context after restarts rather than recomputing it.
GLM-5.2 fused DSA prefill, reported at 845 tok/s versus about 29 tok/s on M3 Ultra, is roughly 30x faster than generic fallback.
It requires macOS 15.0 or later, Python 3.11 to 3.13, and Apple Silicon; multi-Mac inference is explicitly experimental.