v2.6.0
knadh/listmonkv2.6.0Aug 27, 2026by Michael-A-Kuykendall
Upgrading?
AI Summary
A major refactor consolidating the certification pipeline to a 3-box regimen, promoting legacy v2.2 code, and removing legacy engine code to establish Shimmy as a pure Airframe product.
Key Highlights
- Certification pipeline consolidated to 3-box regimen (MATH, INFERENCE, DETERMINISM)
- 26 certified model/quant combinations across 12 families
- Legacy engine code removed (huggingface.rs, mlx.rs, universal.rs)
- CI simplified and cleaned up
- Shimmy is now a pure Airframe product
Breaking Changes
- Removal of engine/huggingface.rs
- Removal of engine/mlx.rs
- Removal of engine/universal.rs
- Removal of src/rustchain_compat.rs
- Removal of RustChain provider documentation and references
New Features
- v2-roadmap.md added
- Chinese user manuals updated
- Safetensors release status clarified
Full Release Notes
## [2.6.0] β 2026-08-26 β Certification Consolidation + Engine Cleanup + v2.2 Promotion ### π Certified β 12 Families Β· 26 Model/Quant Combinations Shimmy's certification pipeline has been **consolidated from the old 5-gate + CHAT regime into a clean 3-box regimen** (MATH + INFERENCE + DETERMINISM), matching the airframe 0.4.0 release. The regression-causing chat box is gone, replaced with deterministic inference verification. **The 3-box certification regimen:** 1. **MATH** β GPU dequant audit, structural peel, plan-vs-peel reds, numerical self-consistency 2. **INFERENCE** β 5-prompt battery (factual / creative / arithmetic / kv_cache / dev_question) on real GPU 3. **DETERMINISM** β two identical runs must produce identical text **26 certified model/quant combinations across 12 families** β up from 25 in v2.4.0. New: llama-3.1-8b, qwen2-7b, gemma-2-2b (plus re-certification of all prior models under the new regimen). ### Highlights - **v2.2 release cleanup promoted to main** β Legacy engine code stripped (huggingface.rs, mlx.rs, universal.rs, rustchain_compat.rs removed β ~1,600 lines deleted). Feature flags cleaned. Shimmy is now a pure Airframe product. - **CI simplified** β Removed huggingface/mlx/llama references from CI/CD. Faster, cleaner builds. - **Certification alignment** β Shimmy's cert documentation, README, and supported-models tables now align with airframe's 3-box regimen and 26-model count. ### Removals - **`engine/huggingface.rs`** (428 lines) β removed - **`engine/mlx.rs`** (323 lines) β removed - **`engine/universal.rs`** (573 lines) β removed - **`src/rustchain_compat.rs`** (306 lines) β removed - **`docs/rustchain-provider.md`** (40 lines) β removed - All RustChain provider references scrubbed from docs, .gitignore, lib.rs ### Documentation - **v2-roadmap.md** added (131 lines) β current roadmap and release planning - Chinese user manuals updated - Safetensors release status clarified ### Infrastructure - **Beads issue tracking initialized** β `.beads/` Dolt-backed board for durable task tracking - **Cargo.lock refreshed** for path-sourced airframe patch ### Commits 18 commits since v2.5.0 β see git log `v2.5.0..HEAD` for full list.
Related Repositories
Get up and running with Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.
A fast reverse proxy to help you expose a local server behind a NAT or firewall to the internet.
Terminal session manager for AI coding agents. One TUI for Claude, Gemini, OpenCode, Codex, and more.
RAGFlow is a leading open-source Retrieval-Augmented Generation (RAG) engine that fuses cutting-edge RAG with Agent capabilities to create a superior context layer for LLMs
The worldβs fastest framework for building websites.