Releases·Model Release
Solo dev releases Lattice: Quark 1.5B trained from scratch and Spark 1.5B Qwen LoRA behind free chat and OpenAI-compatible API
A solo developer shipped Lattice, a pair of small open models served free via chat and an OpenAI-compatible API. Quark 1.5B was trained fully from scratch (random init, custom tokenizer, pretraining plus SFT), while Spark 1.5B is a LoRA fine-tune of Qwen2.5-1.5B-Instruct. Weights are on Hugging Face under MIT and Apache-2.0 licenses respectively, making the full training pipeline reproducible.
5 engagement1 sourceMon, Sep 14, 2026, 08:20 AM
AI Overview
- Solo dev shipped Lattice: Quark 1.5B (MIT, trained from scratch with custom tokenizer, pretraining, SFT) and Spark 1.5B, a LoRA fine-tune of Qwen2.5-1.5B-Instruct (Apache-2.0)
- Both models are free on HF and served with guest chat plus OpenAI-compatible API keys, limited to 20 requests/minute and 500/day; guest chat caps 10 messages/hour per IP
- Dev says the 1.5B single-GPU budget models won't beat Qwen at general knowledge; the goal is an open, reproducible from-scratch pipeline, with a 2.5B from-scratch run planned
Related
Releases·Model Release
Lattice: Quark and Spark, two 1.5B models, now on a free OpenAI-compatible API
Indie developer Lattice made its two 1.5B language models — Quark, trained from scratch, and Spark, a Qwen2.5 fine-tune — available through a free OpenAI-compatible API. The launch matters to practitioners as a zero-cost playground for small experimental models rather than a production-grade offering.
Indie developer Lattice made its two 1.5B language models — Quark, trained from scratch, and Spark, a Qwen2.5 fine-tune — available through a free OpenAI-compatible API. The launch matters to practitioners as a zero-cost playground for small experimental models rather than a production-grade offering.
2 engagement1 sourceredditSun, Oct 11, 2026, 02:49 AM
Products·building
Unified API for open-weight LLMs with free credits during beta
SingularityAPI provides a single API key and base URL to access multiple open-weight LLMs including DeepSeek V4 and Kimi K2. It is OpenAI-compatible, supports streaming, and offers free credits to beta users. The service targets developers who want to experiment with various models without managing separate accounts.
SingularityAPI provides a single API key and base URL to access multiple open-weight LLMs including DeepSeek V4 and Kimi K2. It is OpenAI-compatible, supports streaming, and offers free credits to beta users. The service targets developers who want to experiment with various models without managing separate accounts.
2 engagement1 sourceredditTue, Jul 28, 2026, 09:33 AM
Releases·Model Release
Developer trains 1.1B-parameter LLM from scratch on 20B tokens for about $200
A developer trained a 1.1B-parameter LLM from scratch on 20B tokens from FineWeb-Edu, then fine-tuned it with LoRA on OpenHermes to create a chat model, all for about $200. The project, shared on Reddit, demonstrates a low-cost path to building a custom LLM for learning and resume-building.
A developer trained a 1.1B-parameter LLM from scratch on 20B tokens from FineWeb-Edu, then fine-tuned it with LoRA on OpenHermes to create a chat model, all for about $200. The project, shared on Reddit, demonstrates a low-cost path to building a custom LLM for learning and resume-building.
469 engagement1 sourceredditMon, Aug 10, 2026, 09:44 PM
Products·launched
Chat LLM: no-signup gateway to 300+ models with built-in IDE
Chat LLM is a single interface for comparing 300+ LLMs (GPT, Claude, Llama, DeepSeek, etc.) without signup. It includes a full-stack IDE for building apps in-browser, an image studio, and AI companions with persistent memory. The free tier offers 37 models, and users can bring their own API key.
Chat LLM is a single interface for comparing 300+ LLMs (GPT, Claude, Llama, DeepSeek, etc.) without signup. It includes a full-stack IDE for building apps in-browser, an image studio, and AI companions with persistent memory. The free tier offers 37 models, and users can bring their own API key.
7 engagement1 sourceredditThu, Jul 30, 2026, 06:39 PM
Releases·Tool Release
Quartermaster: open source local AI platform, a llama-swap fork, auto-configures GGUF models behind one OpenAI-compatible API
A developer announced Quartermaster, an open source local AI platform that began as a fork of llama-swap and has since grown into its own project. Users point it at a models folder and it auto-configures runtime settings by reading GGUF headers and measuring free VRAM. It targets ease of use while keeping all settings editable per model.
A developer announced Quartermaster, an open source local AI platform that began as a fork of llama-swap and has since grown into its own project. Users point it at a models folder and it auto-configures runtime settings by reading GGUF headers and measuring free VRAM. It targets ease of use while keeping all settings editable per model.
44 engagement1 sourceredditThu, Sep 3, 2026, 07:36 PM