ds4 and gstack: AI Trend Radar Misreads Mature Open-Source Tools

What happened

An automated open-source trend digest published on October 5, 2026 placed two well-known developer projects near the top of its AI infrastructure category as if they had just appeared. Salvatore "antirez" Sanfilippo's ds4 was listed at 211 stars, all gained that day. Garry Tan's gstack was listed at 125 stars, also all gained that day. 4 In both rows, the total star count matches the daily gain. That pattern usually signals a brand-new repository.

Other reporting tells a different story. ds4 had already collected 8,056 GitHub stars in roughly four days after launch. It also reached the Hacker News front page with 497 points and 157 comments. 2 gstack shows 112,329 stars and 16,695 forks on a skills directory listing. 3 These are not newcomers. They are established tools, and the radar's figures make them look like fresh arrivals.

What ds4 actually is

ds4 is a deliberately narrow inference engine written in C. Its purpose is to run DeepSeek V4 Flash as fast as possible on local hardware. 2 It is not a general GGUF runner and not a wrapper around llama.cpp. 2 The project does credit llama.cpp, GGML and Georgi Gerganov's work as foundations that made it possible. 1

The codebase is split into a few parts: 1

• ds4.c contains the core engine.
• ds4_metal.m handles Metal GPU acceleration.
• A CLI provides an interactive terminal interface.
• An HTTP server speaks both OpenAI- and Anthropic-compatible APIs.
• A test suite compares the logits the engine produces.

The headline benchmark is 26.68 tokens per second of generation for the 284-billion-parameter Mixture-of-Experts model. That figure comes from a 128 GB MacBook Pro with an M3 Max. The engine also supports a 1-million-token context window and a persistent on-disk KV cache. 2 Coverage also mentions asymmetric Q2 quantization. 1

A roadmap tracker shows the project kept evolving after launch. The work included: 1

• CUDA refactoring
• iMatrix quantizations
• DGX Spark support
• New model targets in mid-September, DeepSeek 4.1 Flash and Qwen3.8 Flash Next

The same tracker reports that the main branch went quiet between September 20 and October 4, while pull requests kept arriving. 1 The radar's one-line summary describes ds4 as covering DeepSeek 4 Flash and PRO across Metal, CUDA and ROCm. 4 That is broader than what other coverage details, so the ROCm and PRO claims are worth treating cautiously until verified against the repository itself.

What gstack actually is

gstack packages Garry Tan's own Claude Code setup as 23 opinionated tools. Each tool plays a role, such as CEO, designer, engineering manager, release manager, documentation engineer or QA. 3 It is written mainly in TypeScript and ships a SKILL.md manifest so compatible agents can discover it. Its listing says it works with Claude Code, Codex CLI and ChatGPT. 3 The radar describes it in nearly the same terms, as a plug-and-play agent orchestration kit. 4 The real disagreement is about scale, not substance.

Why the numbers diverge

The most likely explanation is a data artifact, though this is an inference rather than a confirmed cause. When a tool reports a repository's total stars as exactly equal to its daily gain, it suggests the pipeline is only counting stars seen in its own observation window. That can happen if it reads from a trending feed or a limited event stream instead of the repository's full metadata.

The same day's report has the same pattern elsewhere. For example, calesthio/OpenMontage appears at 245 total and +245 today. 4 Read correctly, the daily numbers can still be useful, since they show continued momentum for ds4 and gstack weeks after their debuts. The "total" column, however, cannot be taken at face value.

Why it matters

Automated radars are becoming a common way developers, investors and journalists scan the AI tooling landscape. The October 5 digest frames its day around several trends: 4

• agent memory
• web access for agents
• retrieval tools
• local-first runners such as Ollama and Open WebUI

Those broad themes may be sound. But a digest that presents a 112,000-star project as a 125-star debut misleads anyone using it to spot emerging work. It hides which projects are genuinely new and which are mature tools still drawing attention.

The practical lesson is simple. Treat generated trend reports as pointers, not measurements. Check star counts, release history and commit activity at the source before drawing conclusions about adoption.

On the substance, both projects are worth attention. ds4 stands out as a focused bet on running a frontier-scale model on a laptop. gstack is a widely forked template for role-based agent workflows. Neither is new, and the more interesting story is how firmly established both already are.