Tag

club-3090

Provides reusable Docker compose files, scripts and benchmarked configs to serve modern LLMs (Qwen, Gemma, etc.) on 1–2 NVIDIA RTX 3090/4090/5090 GPUs. Multi-engine (vLLM, llama.cpp, ik_llama), measured TPS/context tradeoffs, and validated single/dual‑GPU recipes.

MTPLX

Runs local LLMs on Apple Silicon using native MTP speculative decoding to accelerate token generation while preserving the model's output distribution. Leverages the model's own MTP heads with batched verification and exact rejection sampling; ships with a Mac app, CLI, local OpenAI/Anthropic-compatible server, auto-tune, and Forge for building/verifying MTP adapters.

BugTraceAI-CORE-Ultra-27B-Q6

Generates production-ready offensive-security artifacts from prompts—Nuclei templates, CVE PoCs, exploit scripts and pentest tooling—fine-tuned on bug-bounty reports and CVE writeups and quantized for consumer/server GPU deployment.

MOSS-Transcribe-Diarize

Converts long-form multi-speaker audio/video into a compact, speaker-aware transcript with timestamps and anonymous speaker labels in one pass. Combines ASR and diarization in a single model, supports custom prompts/hotwords, and targets meetings, podcasts, interviews and long recordings.

Fara1.5-27B

Automates end-to-end web workflows from browser screenshots by emitting pixel-grounded actions (click, type, scroll, visit, search). Vision-first multimodal agent fine-tuned from Qwen3.5-27B with critical-point safety checks; intended for sandboxed, human-supervised deployments.

Munder Difflin

Turns terminal-agent CLIs you already run into a local desktop multi-agent harness: each agent runs as a real terminal process, with shared semantic memory, encrypted on-node messaging, a GOD orchestrator for routing/approvals, and a visual office floor for monitoring.

lordx64/agentic-distill-fable-5-sft

Provides 4,659 agentic single-turn SFT training pairs extracted from Claude Fable‑5, formatted as a single-column parquet for Qwen-style fine-tuning. Includes explicit chain-of-thought (<think>) blocks, XML-serialized <tool_use> calls, PII redaction, and AGPL-3.0 licensing.

Qwable-v1

Open-weights agentic coding model that layers Claude Fable‑5 tool‑use SFT onto a reasoning‑distilled Qwen3.6 base; emits <tool_use> XML for file edits, shell commands and reads when prompted as an agent. Designed for agentic coding workflows; AGPL‑3.0 licensed.

Qwable 27B

Fine-tuned full checkpoint of the Qwen3.6-27B base that produces structured, trace-style assistant outputs for code, technical reasoning, and instruction-following. Packaged for local GGUF conversion and local inference; not a LoRA adapter and not validated for production use.

Krea 2 (Comfy-Org/Krea-2)

Provides ComfyUI-ready repackaged checkpoints of the Krea 2 image model family for local text-to-image workflows. Includes RAW (undistilled base for fine-tuning and LoRA training) and Turbo (8-step distilled checkpoint for fast inference), using a Qwen Image VAE and Qwen3‑VL encoder.

Zone of Proximal Policy Optimization: Teacher in Prompts, Not Gradients

Proposes ZPPO, a distillation method that keeps the teacher inside prompts rather than injecting teacher gradients, using binary- and negative-candidate prompts plus a prompt replay buffer to recover learning signal on hard examples; shows gains for small Qwen3.5 students across 31 multimodal benchmarks.

Beyond the Current Observation: Evaluating Multimodal Large Language Models in Controllable Non-Markov Games

Evaluates multimodal LLMs' ability to reconstruct past observations and act in controllable non-Markov games. Introduces RNG-Bench with two games (Matching Pairs, 3D Maze), three controllable difficulty axes, a head-to-head duel protocol, and a Memory Gap metric to separate forgetting from action errors.