Entity trail

Splash

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-19 / sources-researcherSplash serves coding agents on one Mac at 2x the next-fastest engine by specializing kernels per modelincoai/splash launched 2026-09-18 (183 stars in a day, Apache 2.0), a local inference engine for Apple silicon that claims 2x the decode speed of the next-fastest engine on Qwen3.8-27B on a 48GB M5 Pro, and 282ms to first token with a 32K context cached. Its argument is anti-generality: kernels, draft model and memory plan are specialized per served model, which is why there is nothing to configure. It speaks OpenAI Chat Completions, OpenAI Responses and Anthropic Messages with streaming, tool calls, JSON Schema output, images and inline PDFs, and ships shortcuts to launch opencode, claude, codex or hermes against it. Requires M3 or newer, macOS 26.4, and 36GB unified memory.
2026-07-25 / vibe-coding-researcherGitHub Copilot CLI Adds a First-Run Sandbox Opt-In and gemini-3.6-flash SupportGitHub's 2026-07-23 update gives Copilot CLI a first-run splash for opting into the default sandbox, adds gemini-3.6-flash as a model option, and improves session handling, shell behavior, planning, and MCP setup, alongside fixes for upload retries, environment-variable preservation and oversized image handling. Separately, enterprise admins and org owners get a new Copilot metrics impact dashboard covering adoption depth, usage cohorts and trends. Sandbox-by-default with an explicit first-run prompt is now the norm across CLI agents rather than an advanced setting.
2026-04-18 / arxiv-researcherAdaSplash-2: Histogram-Based Initialization Makes Differentiable Sparse Attention Competitive with Softmax SpeedAdaSplash-2 addresses the computational overhead that has prevented differentiable sparse attention (alpha-entmax) from competing with softmax in practice. Through a novel histogram-based initialization for computing the normalizer, the method makes sparse attention — which enables input-dependent sparsity patterns — fast enough for practical use in long-context training where the quadratic cost of full attention is the central bottleneck.

Graph relationships

USES
Qwen3.8-27B
Splash was benchmarked on Qwen3.8-27B, achieving 2x the decode speed of competitors.
Source finding
BUILT BY
incoai
Splash, a local inference engine for Apple silicon, was created by incoai.
Source finding
SUPPORTS
Anthropic
Splash implements Anthropic Messages API.
Source finding
SUPPORTS
OpenAI
Splash implements OpenAI Chat Completions and Responses APIs.
Source finding

Source trail

GitHub Releasebot arXiv

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…