Tiny Decision Encoders Go Local, Plus 280+ MCP Servers and Hybrid RAG Builds

Today's findings

1. #1 DecisionTune 1.0: 395M Local Encoder for Fast Option Pickingrepo

   A 395M Apache-2.0 encoder picks among options or answers yes/no in about 10 ms on MLX, giving agents a cheap local decision layer.

   DecisionTune 1.0 · decision layer

   Most agent decisions don't need a generative model

   vs

   LLM call

   395M encoder

   Latency

   Round-trip call

   ~10 ms on MLX

   Cost

   Billed tokens

   Local compute

   Data

   Leaves device

   Stays on-device

   Output

   Free-form text

   Option pick / yes-no

   LLM call wins the row 395M encoder wins the row

   Routing, triage, and quality gates are classification problems — swap them to the encoder, keep the LLM for generation,

   Why it matters: Most agent routing and triage decisions do not need a generative LLM; a small local encoder can cut latency and cost while keeping data on-device.

   How to apply: Run it on MLX or export to ONNX, then replace LLM calls for tool selection, ticket routing, and answer-quality gates; benchmark against your current router.

   local-llmagentsroutingmlx

   Read more: DecisionTune 1.0: a 395M encoder that picks from your options offline, about 10 ms per short decision on MLX (Apache-2.0)

2. #2 Curated List of 280+ Official MCP Serversrepo

   A maintained GitHub list collects official MCP servers and actively maintained open-source projects, grouped by category.

   MCP ECOSYSTEM

   One maintained list, every official MCP server

   280+

   official & actively maintained MCP servers

   curated, not scattered

   Official

   repos only — avoids abandoned forks

   5 areas

   databases, cloud, observability, search, design

   claude mcp add

   pin the official repo to install

   Official repos only — fewer abandoned forks, lower supply-chain risk.

   Why it matters: MCP integrations are easy to add but hard to trust; using official repos avoids abandoned forks and reduces supply-chain risk.

   How to apply: Browse the list for databases, cloud, observability, search, and design servers; add with `claude mcp add` and pin the official repo.

   mcpclaudetooling

   Read more: Official MCP servers, updated list (280+)

3. #3 Hybrid RAG Pipeline with Dense Search, BM25, RRF, and Rerankingtechnique

   A full walkthrough and repo build a hybrid retrieval pipeline and test it against vector-only and BM25 baselines.

   Why it matters: Hybrid retrieval plus reciprocal rank fusion and reranking is a strong, practical baseline for production RAG when pure vector search misses exact terms.

   How to apply: Clone the repo, run the ReRankEval harness on your corpus, and compare dense-only, BM25-only, RRF, and LLM reranking before tuning embeddings.

   ragretrievalreranking

   Read more: Building and Testing a Hybrid RAG Pipeline — Dense Search, BM25, RRF, and Reranking (Part 1) · Hybrid RAG Pipeline — Dense Search, BM25, RRF, and Reranking (Part 1) · Hybrid RAG Pipeline — Dense Search, BM25, RRF, and Reranking (Part 1) · Building and Testing a Hybrid RAG Pipeline — Dense Search, BM25, RRF, and Reranking (Part 1)

4. #4 Strata Runs Qwen3.8-Flash-Next Locally on Prosumer GPUs and Macstechnique

   Multiple reports show Strata with NVFP4/IQ3 quantized Qwen3.8-Flash-Next hitting 60-105 tok/s on RTX/V100 rigs and 17.5 tok/s on a 64 GB Mac mini with SSD streaming.

   Why it matters: Large MoE models are becoming practical on local hardware when paired with aggressive quantization, expert caching, and SSD offload.

   How to apply: Try the Strata NVFP4 fork with Qwen Flash Next quantized weights; tune expert cache size, KV cache, and SSD streaming for your VRAM/RAM budget.

   local-llmquantizationmoestrata

   Read more: ~188k warm ~60–67 tok/s: Qwen3.8-Flash-Next NVFP4 with Strata on a single RTX PRO 4500 32GB + 64GB DDR5. · ~188k warm ~60–67 tok/s: Qwen3.8-Flash-Next NVFP4 with Strata on a single RTX PRO 4500 32GB + 64GB DDR5. · Qwen3.8-Flash-Next-IQ3_S on Strata V100 @130 Watts Results. · 4x 3080 20GB (modded, alibaba) + Strata (Qwen Flash Next 125B IQ3_XXS) = 105 t/s generation 5000 prompt processing on 150k context · I got Qwen Flash Next Q4 running on a Mac Mini m5 64gb with ssd streaming · Got Qwen Flash Next Q4 running on my Mac Mini M5 64GB with ssd streaming · MoE SSD streaming on a 64 GB Mac mini: GPU still waits 27% of decode on experts. Ideas? · I am really geeking out - Strata + Qwen 125bn

5. #5 Aleph Alpha Releases Kolibri 78B Open-Weight MoErepo

   Kolibri is a 78B-parameter, 3.46B-active open-weight model with up to 1M tokens of context and an Apache-2.0 license.

   architecture

   78B capacity, 3.46B firing per token

   router

   top-4 gate

   total parametersactive per token

   weighted

   merge

   78B

   total params

   3.46B

   active per token

   ~22x

   sparsity ratio

   1M

   max context tokens

   Why it matters: It adds a sovereign, permissively licensed long-context MoE option for teams that need on-prem or EU-hosted deployments.

   How to apply: Pull the Hugging Face weights, quantize for your hardware, and evaluate long-document summarization, retrieval, and multilingual tasks against Qwen/Llama baselines.

   open-weightsmoelong-contextlocal-llm

   Read more: German lab Aleph Alpha releases Kolibri: a sovereign open-weight model,78B parameters, 3.46B active. Up to 1M tokens of context. · You can now try Aleph Alpha's Kolibri 78B for free online here. · You can now try Aleph Alpha's Kolibri 1 78B for free online here.

6. #6 Git Worktrees for Parallel Claude Code Sessionstip

   A guide explains how to isolate multiple Claude Code sessions with git worktrees and avoid path drift and shared-state corruption.

   PARALLEL CODING AGENTS

   One shared directory — or one worktree per session

   Shared working directory

   vs

   Worktree per session

   Isolate each agent session in its own worktree before running in parallel.

   Same repo, separate checkouts — parallel agents never share state to corrupt.

   Why it matters: Running parallel coding agents in one working directory can corrupt the Git index and create race conditions; worktrees are a simple isolation fix.

   How to apply: Create one worktree per agent session (`git worktree add ../feature-branch feature-branch`), launch Claude Code inside it, and keep CWD scoped to that worktree.

   claudegitagents

   Read more: The Git Worktree guide for parallel Claude Code sessions: fixing path drift and shared state corruption

7. #7 Alter Zero: Rust Terminal Agent Harness with Live Sessionsrepo

   Alter Zero is an open-source, RAM-efficient terminal agent harness that interacts with live terminal sessions instead of one-shot shell commands.

   Why it matters: Live terminal control is useful for coding, cybersecurity, and automation tasks where state and interactive prompts matter.

   How to apply: Run it with your preferred local or Claude model, wire it to a sandboxed terminal, and use it for multi-step shell workflows that need persistent context.

   agentsrustterminal

   Read more: Alter Zero: open-source RAM efficient terminal agent harness for coding, cybersecurity, and automation Built in Rust.

8. #8 Kaoru: Open-Source Desktop Agent with Memory and Permissionsrepo

   Kaoru is a free, open-source desktop AI agent with memory, tools, and permission controls.

   Why it matters: Permission controls and persistent memory are two missing pieces in many desktop agent setups; Kaoru provides a reference implementation.

   How to apply: Install it locally, configure tool permissions, and inspect its memory/permission model before building your own desktop agent.

   agentslocalpermissions

   Read more: I built Kaoru, a desktop AI agent with memory, tools and permission controls — it’s free and open source if you want to try it

9. #9 LangGraph + MCP Behind Next.js: Fix SSE Buffering and Gate Email Sendstip

   A build note covers two production gotchas: Next.js rewrites buffer SSE chunks, and email-sending tools need explicit human gating.

   LangGraph + MCP, behind Next.js

   Two production gotchas with this setup

   SSE buffering Next.js rewrites hold stream chunks — disable buffering on SSE routes or the stream looks broken

   Ungated sends Email tools must wait for an explicit human approval gate before executing

   Reliability gotcha Safety gotcha

   Both fail by default; both have one-line fixes.

   Why it matters: Streaming agents behind a web framework often appear broken due to buffering, and write tools like email need approval before execution.

   How to apply: Disable buffering for SSE routes and add an approval gate before any email-send tool call; test with a LangGraph + MCP agent behind Next.js.

   mcplanggraphnextjsagents

   Read more: Two things that cost me time building a LangGraph + MCP agent behind Next.js: SSE buffering and gating email sends

10. #10 Ornith 1.5 35B-A3B Triples Local Agent Throughputtip

    A local agent switched from Qwen3.8 27B to Ornith 1.5 35B-A3B on two RTX 5070 Tis, getting about 180 tok/s vs 60 with the same scores on two agent tests.

    Why it matters: Model swaps can deliver large speed gains without losing task quality, making local agent loops more practical.

    How to apply: Test Ornith 1.5 35B-A3B in Ollama on your hardware with a long-session tool-call test and a small coding suite; compare against your current local model before switching.

    local-llmagentsollama

    Read more: Switched my local agent from Qwen3.8 27B to Ornith 1.5 35B-A3B on two 5070 Tis: about 180 tok/s vs 60, same scores on my tests

11. #11 Distilling Stockfish on a Billion Positions with a 3.9B Datasetpaper

    A project distills Stockfish's value function into ResNet/ViT models using 1B positions and releases a 3.9B-position dataset.

    Why it matters: It provides a large open dataset and a reproducible distillation setup for chess evaluation, useful for anyone studying value models or distillation at scale.

    How to apply: Download the dataset, train a ResNet/ViT value model, and use the released pipeline to benchmark distillation techniques on your own domain.

    distillationdatasetchess

    Read more: Distilling Stockfish on a Billion Positions, Full 3.9B Dataset Available [P]

12. #12 UCVG.cpp Generates Control Vectors for Any LLM in C++repo

    UCVG.cpp is a C++ tool that generates control vectors for any LLM from a single prompt pair.

    UCVG.cpp · control vectors

    One prompt pair steers any LLM — no fine-tuning

    1

    Prompt pair

    one positive, one negative

    the entire input — no training data needed

    2

    Control vector

    generated from the pair

    3

    Inference

    applied during the run

    4

    Steered output

    tone · refusal · task focus

    C++ implementation fits llama.cpp-style local stacks.

    Why it matters: Control vectors let you steer model behavior without fine-tuning, and a C++ implementation fits local/llama.cpp-style stacks.

    How to apply: Generate a control vector from a positive/negative prompt pair, then apply it during inference to steer tone, refusal, or task focus.

    control-vectorscpplocal-llm

    Read more: Control vector generation tool in C++ for any LLM in a single prompt pair. (UCVG.cpp)