Having fun with oh my pi, DeepSeek-V4-Flash, GPT-5.6 Luna and Antigravity CLI

7 hours ago

The author describes a setup using the coding agent 'omp', which is built on the minimal 'pi' agent and offers a batteries-included experience similar to Codex and Claude Code.
DeepSeek-V4-Flash, now out of preview, is a cheap and powerful model that lacks multimodal capabilities; the author uses GPT-5.6 Luna Max via omp to provide vision functionality.
To enable web search in omp, the author created a custom extension that uses Antigravity CLI (agy) for search instead of native search providers like Exa or Perplexity.
The final setup combines DeepSeek-V4-Flash for coding, GPT-5.6 Luna Max for vision, and agy CLI for search, all integrated through omp.

DeepSeek-V4-Flash-0731 can be used with Transformers via pipeline or loading model directly, with code examples provided.
It is also deployable using vLLM (pip install or Docker) with DSpark speculative decoding enabled via a specific flag.
SGLang serves the model via pip or Docker, and the Docker Model Runner command is provided.

The author strongly desires local models to be competitive with hosted APIs for practical use in coding agents, but the current experience is fragmented and overly complex.
Key frustration: local model setup requires many choices (inference engine, model, quantization, template, context size) and config tweaks, often leading to poor results or broken functionality.
Tool parameter streaming is a critical missing feature in local models—without it, users face long dead connections, can't see tool calls as they happen, and lose ability to interrupt early.

<antirez>6 days ago

DwarfStar 4 (DS4) became popular quickly due to the release of a quasi-frontier model that is large and fast enough for local inference, combined with an asymmetric 2/8 bit quant setup that works on 96-128GB RAM.
The project leverages the experience of the local AI movement and GPT 5.5, allowing rapid development (one week) with proper LLM interaction skills.
The author worked 14 hours/day on average during the last week, similar to the early Redis days, but normal average is 4-6 hours.

grep is a ubiquitous text search tool in Linux environments.
The guide covers basics, recursive search, search options, and output options with interactive examples.
Basic usage involves providing a pattern and file; grep prints matching lines.

DeepSeek-V4-Flash-0731 can be used with Transformers via pipeline or loading model directly, with code examples provided.
It is also deployable using vLLM (pip install or Docker) with DSpark speculative decoding enabled via a specific flag.
SGLang serves the model via pip or Docker, and the Docker Model Runner command is provided.

The author strongly desires local models to be competitive with hosted APIs for practical use in coding agents, but the current experience is fragmented and overly complex.
Key frustration: local model setup requires many choices (inference engine, model, quantization, template, context size) and config tweaks, often leading to poor results or broken functionality.
Tool parameter streaming is a critical missing feature in local models—without it, users face long dead connections, can't see tool calls as they happen, and lose ability to interrupt early.

<antirez>6 days ago

DwarfStar 4 (DS4) became popular quickly due to the release of a quasi-frontier model that is large and fast enough for local inference, combined with an asymmetric 2/8 bit quant setup that works on 96-128GB RAM.
The project leverages the experience of the local AI movement and GPT 5.5, allowing rapid development (one week) with proper LLM interaction skills.
The author worked 14 hours/day on average during the last week, similar to the early Redis days, but normal average is 4-6 hours.

grep is a ubiquitous text search tool in Linux environments.
The guide covers basics, recursive search, search options, and output options with interactive examples.
Basic usage involves providing a pattern and file; grep prints matching lines.

Hasty Briefsbeta