Skip to content

MIRA v0.318.7 — public beta

Pre-release
Pre-release

Choose a tag to compare

View all tags
@tarekedOz tarekedOz released this 27 Jul 12:01
v0.318.7
6c03fdd

MIRA v0.318.7 — public beta

MIRA is a self-hostable AI assistant that runs as a single binary. Published by Vexillon under AGPL-3.0-or-later.

Public beta. Supersedes v0.318.0. A stabilization + context-efficiency series (0.318.1–0.318.7) focused on running reliably against local models. No new configuration and no breaking changes — safe to upgrade in place.

Leaner prompts on local models

  • Up to ~83% fewer prompt tokens per turn. With prompt caching and adaptive tool selection on, MIRA now ships a small, stable core toolset plus on-demand tool search instead of every tool's schema on every turn. In a measured chat turn this dropped the prompt from ~24K to ~4K tokens — cheaper, faster, and far easier for a small local model to handle.
  • Context budgeted to what the model can actually use. MIRA sizes the working window to a model's reliable-recall range rather than an inflated advertised maximum, keeps injected memory/wiki context to a modest share of that window (prioritizing curated wiki), and now includes the tool-schema block in its per-turn token estimate (previously undercounted by ~10x).
  • Startup context sanity check. On start, MIRA best-effort checks your configured context length against the model's actually-loaded window — for both LM Studio and Ollama — and warns when they don't match (a common cause of silent truncation with Ollama's small default context).

Reliability & correctness

  • Degeneracy guard now also watches the reasoning stream. A reasoning model that gets stuck repeating itself while "thinking" is now caught and the turn is failed early, instead of running to the token cap.
  • No more HTTP 400 from stacked system messages. MIRA coalesces system content into a single leading block before sending, so local chat templates that reject multiple or non-leading system messages work correctly.
  • Config saves are atomic. Writes go through a temp file with fsync + rename and a backup, so an interrupted or racing save can no longer truncate your live config.
  • Large tool results can't overflow a small model — they're clamped to the effective window mid-conversation.
  • Per-turn model overrides are covered by the degeneracy guard (previously they bypassed it).
  • UI-set provider keys win over a stale .env. A key set in the app is no longer silently overridden by an environment value; .env now only fills an empty setting, with a loud warning when it would shadow a configured key.

Install

# Linux / macOS
curl -fsSL https://get.vexillon.ai/install.sh | sh
# Windows
irm https://get.vexillon.ai/install.ps1 | iex

Targets

  • x86_64-unknown-linux-gnu
  • aarch64-unknown-linux-gnu
  • x86_64-apple-darwin
  • aarch64-apple-darwin
  • x86_64-pc-windows-msvc

Verify

All artifacts are signed with minisign. Verify the checksum file, then the archive checksums:

minisign -Vm SHA256SUMS -P RWQX8PcjZB3S8i1V9AkZgUKoIy+eZWRzaEHKvBW9U/mkMP9Vp0B8xOaO
sha256sum -c SHA256SUMS

Trusted comment on a valid signature reads: signed by vexillon.ai — MIRA 0.318.7

Assets 14
Loading