Erik Kristensen and Napalys Klicius write about four changes to the GitHub Copilot harness that reduce token costs without sacrificing task quality. The central insight is that optimizing individual tool calls is the wrong metric — a shorter response can cost more overall if it forces the agent to rerun commands or reread output. The four changes are: selectively compressing repetitive build/test/install output while preserving source-like content, removing unused line-number prefixes from file reads, halving the task-tool prompt via a meta-prompting loop, and batching background completion notifications so results arrive without an extra retrieval turn. Each was validated through offline agentic benchmarks and controlled online A/B experiments before shipping.
- RTK (Rust Token Killer) was evaluated and found to increase end-to-end cost despite shortening individual responses, because the agent reopened or reran commands to recover omitted details.
- The prompt compression initially caused a regression that offline tests missed: cautious parallelism guidance was rewritten into a hard scheduling policy, serializing independent agents. The fix was a single sentence: "Independent agents can run in parallel; consider side effects."
- A tighter file-tool instruction set that worked in Copilot code review actually increased cost in Copilot CLI, illustrating that evidence is local to the workload.
- The changes ship across all Copilot products sharing the same harness (CLI, app, code review); code review separately saw ~20% cost reduction from a prior migration to shared file tools.
- RTK (Rust Token Killer) was evaluated and found to increase end-to-end cost despite shortening individual responses, because the agent reopened or reran commands to recover omitted details.
- The prompt compression initially caused a regression that offline tests missed: cautious parallelism guidance was rewritten into a hard scheduling policy, serializing independent agents. The fix was a single sentence: "Independent agents can run in parallel; consider side effects."
- A tighter file-tool instruction set that worked in Copilot code review actually increased cost in Copilot CLI, illustrating that evidence is local to the workload.
- The changes ship across all Copilot products sharing the same harness (CLI, app, code review); code review separately saw ~20% cost reduction from a prior migration to shared file tools.
Brendan McKeag writes about Qwen3.8-27B, a dense 27B-parameter vision-language model from Alibaba that achieves frontier-class agentic coding performance while fitting on a single 24GB GPU. The model's hybrid attention stack (48 linear-attention layers against 16 full attention layers) slashes KV cache memory, and its dense architecture makes it far more practical for fine-tuning than the MoE alternatives dominating the industry. McKeag argues the model serves best as an active tool-calling assistant rather than an encyclopedia, and its small footprint makes it uniquely suited for Runpod Serverless scale-to-zero deployments where cold start is dominated by bytes on disk.
- The model is roughly 3x more token-hungry than Qwen3.6-27B, trading wall-clock speed for better answers
- It crossed 1M Hugging Face downloads within days of release; the community nickname is "local Opus"
- Unsloth shipped QLoRA support the day after weights dropped; 4-bit QLoRA is the well-trodden dense route, while MoE QLoRA at 4-bit is explicitly unsupported
- FlashBoot snapshots require loading the model at worker boot before runpod.serverless.start() to be effective
- Harvey and Engram reported it averaging 67% across 250 legal tasks in a synthetic law firm of ~10,000 documents, ahead of every model in that study
- The built-in multi-token prediction head enables speculative decoding without a separate draft model
- The model is roughly 3x more token-hungry than Qwen3.6-27B, trading wall-clock speed for better answers
- It crossed 1M Hugging Face downloads within days of release; the community nickname is "local Opus"
- Unsloth shipped QLoRA support the day after weights dropped; 4-bit QLoRA is the well-trodden dense route, while MoE QLoRA at 4-bit is explicitly unsupported
- FlashBoot snapshots require loading the model at worker boot before runpod.serverless.start() to be effective
- Harvey and Engram reported it averaging 67% across 250 legal tasks in a synthetic law firm of ~10,000 documents, ahead of every model in that study
- The built-in multi-token prediction head enables speculative decoding without a separate draft model
Joe Rice-Jones writes that his Asus ROG Flow Z13 tablet with a Ryzen AI Max+ 395 APU and 128 GB unified memory now runs the same local LLMs as his desktop mini PCs via Lemonade, using a model router to automatically send prompts to different sized models for everyday use and agentic coding with Crush, making it his preferred device.
- Tablet is the Kojima Productions Ludens special edition ROG Flow Z13
- Radeon 8060S iGPU has 96 GB of memory allocated in Adrenalin software
- Model tiers include Qwen3.5-2B to gpt-oss-120b MXFP4 ~59 GB on disk
- 120B model loads in 27.6 s, first token in 0.86 s, ~52 tokens/s on iGPU
- Misrouted CPU-only run dropped to 0.44 tokens/s, over 100x slower
- Router uses Qwen3.5-0.8B classifier and logs which rule fired per response
- Crush coding required raising Lemonade context window from 4,096 to 32,768 tokens
- Tablet is the Kojima Productions Ludens special edition ROG Flow Z13
- Radeon 8060S iGPU has 96 GB of memory allocated in Adrenalin software
- Model tiers include Qwen3.5-2B to gpt-oss-120b MXFP4 ~59 GB on disk
- 120B model loads in 27.6 s, first token in 0.86 s, ~52 tokens/s on iGPU
- Misrouted CPU-only run dropped to 0.44 tokens/s, over 100x slower
- Router uses Qwen3.5-0.8B classifier and logs which rule fired per response
- Crush coding required raising Lemonade context window from 4,096 to 32,768 tokens
This guide provides a comprehensive setup for using the Ghostty terminal emulator with Anthropic's Claude Code agentic coding tool. It highlights how Ghostty’s native splits, notification forwarding, and state restoration features minimize friction during long sessions where an AI agent is modifying files and running tests.
The article covers several key optimizations:
- Core configuration for typography (JetBrains Mono Nerd Font), themes, and window persistence to restore layouts between restarts.
- Implementation of the SAND keybinding pattern for intuitive split pane management.
- Three distinct workspace presets: Standard Split, Three-Pane Neovim Layout, and Multi-Project Tab Layout.
- Automation scripts for installing necessary fonts, configuring a compatible Starship prompt (to avoid issues with Powerlevel10k), and setting Ghostty as the default terminal handler.
- A specialized tmux script designed to handle remote sessions while ensuring Claude Code notifications and extended keys pass through correctly.
- Specific JSON settings to synchronize Claude Code's theme and system notification behavior with Ghostty’s environment.
The article covers several key optimizations:
- Core configuration for typography (JetBrains Mono Nerd Font), themes, and window persistence to restore layouts between restarts.
- Implementation of the SAND keybinding pattern for intuitive split pane management.
- Three distinct workspace presets: Standard Split, Three-Pane Neovim Layout, and Multi-Project Tab Layout.
- Automation scripts for installing necessary fonts, configuring a compatible Starship prompt (to avoid issues with Powerlevel10k), and setting Ghostty as the default terminal handler.
- A specialized tmux script designed to handle remote sessions while ensuring Claude Code notifications and extended keys pass through correctly.
- Specific JSON settings to synchronize Claude Code's theme and system notification behavior with Ghostty’s environment.
This repository provides the GGUF quantized weights for Qwen3.6-27B, a flagship-level coding model designed for stability and real-world utility. The model features significant upgrades in agentic coding capabilities, allowing it to handle frontend workflows and repository-level reasoning with high precision. It also introduces thinking preservation, which enables the model to retain reasoning context from historical messages to improve iterative development.
Key technical highlights:
* Native context length of 262,144 tokens, extensible up to 1,010,000 via RoPE scaling (YaRN).
* Enhanced tool-calling capabilities for complex agentic tasks.
* Support for multimodal inputs including images and video.
* Optimized for various inference frameworks like SGLang, vLLM, and KTransformers.
Key technical highlights:
* Native context length of 262,144 tokens, extensible up to 1,010,000 via RoPE scaling (YaRN).
* Enhanced tool-calling capabilities for complex agentic tasks.
* Support for multimodal inputs including images and video.
* Optimized for various inference frameworks like SGLang, vLLM, and KTransformers.
