Strim 0.1.0

Strim is a Swift-only local inference engine and agent backend for Qwen3.8-27B on Apple Silicon: one binary that serves an Anthropic-compatible API for Claude Code and Codex. MIT licensed.

Measured against rapid-mlx and mlx-vlm (Qwen3.8-27B 4-bit, one 64 GB M5 Max, medians of 2 passes):

• Decode at 4K context: 70.72 tok/s vs 47.75 (rapid-mlx) and 30.69 (mlx-vlm).
• Four agents sharing a 16K prefix: median first token 20.71 s vs 75.08 s and 70.93 s.
• 30-55% lower median first token at 4 to 8 concurrent requests.

Where it is not faster:

• Cold 32K prefill: first token 38.04 s vs 36.61 s and 36.25 s (about 5% behind the best rival).
• Aggregate throughput at 4 to 8 concurrent requests trails the best rival by more than 2% in 3 of 6 cells (5-6%); at 8 agents mlx-vlm is ahead (30.78 vs 28.94 tok/s).
• p95 first-token latency is more than 2% worse in 2 of 6 cells (5-7%).

Preliminary benchmark caveat: all engine sessions in this run were measured while other jobs loaded the machine (highest 1-minute load 11.8). A clean rerun is in progress. Only one machine was tested.

Also in this release: structured output, on-disk prefix cache, scoped agent benchmark, strim pull (resumable, verified model download), strim doctor.

Install:
brew install megasoft1978/strim/strim
# or
curl -fsSL https://raw.githubusercontent.com/megasoft1978/strim-engine/main/packaging/install.sh | bash

Requirements: Apple Silicon Mac with 64 GB memory, macOS 14 or later. The binary is ad-hoc signed (no Developer ID, not notarized); Homebrew and curl installs are not quarantined, a tarball downloaded in a browser may be blocked by Gatekeeper. Model weights (about 15 GB) are downloaded separately with strim pull. Details: docs/INSTALL.md.