Global

A high-throughput and memory-efficient inference and serving engine for LLMs

Implementing LLM architectures from scratch in PyTorch

sh as its gate, so latest is always the current stack. /models), and serves on port 18020. Pick a mode — one GPU serves one at a time:

Jev-like family of decision models built on top of Qwen3.5/3.8 you can train and run on your own

Qwen3-TTS is an open-source series of TTS models developed by the Qwen team at Alibaba Cloud, supporting stable, expressive, and streaming speech generation, free-form voice design, and vivid voice cloning.

This release contains training, two inference backends, benchmark scoring, temperature calibration, and a browser demo.

8-27B bf16 (130 tok/s), 262k context, prefix caching. Works with Claude Code, Codex, opencode and pi.

《开源大模型食用指南》针对中国宝宝量身打造的基于Linux环境快速微调（全参数/Lora）、部署国内外开源大模型（LLM）/多模态大模型（MLLM）教程

🔥 MaxKB is an open-source platform for building enterprise-grade agents. 强大易用的开源企业级智能体平台。

🍔 In addition, ms-swift integrates the latest training technologies, including Megatron parallelism techniques such as TP, PP, CP, EP to accelerate training, as well as numerous GRPO algorithm family reinforcement learning algorithms including: GRPO, DAPO, GSPO, SAPO, CISPO…

The fastest way to run Qwen3.8-Flash-Next on Strix Halo (gfx1151)

W&B Training (Serverless RL) is the first publicly available service for flexibly training models with reinforcement learning.

> New project: JevHarness — Let an LLM build task-specific decision harnesses with Jev, with optional refinement using rewards and execution traces. Includes an interactive Pokémon demo.

app — the end-signal probe, public pages, and the live steered-model lab.

What it is, and what it isn't. These are very small models. They make extremely fast, well-calibrated judgement calls between options, and they slot easily into your local code.

Qwen3 is the large language model series developed by Qwen team, Alibaba Cloud.

8-Flash-Next is the foundation model developed by Qwen Team, Alibaba Group.

Cross-platform: Mac M5 Pro (Q3KM · 13.5GB · Metal) + Windows (Q4KM · 16.8GB · CUDA).

This is wire-compatible, but not mathematically equivalent to OpenJev's logit read. The probabilities are generated/self-reported by the model rather than read directly from its logits.

Run a 105 GB AI model on a Mac that can't hold it. 8-Flash-Next (125B mixture of experts) from your SSD and caches the busiest experts in memory, so it runs on Macs with 16 to 64 GB. One native Swift binary on MLX and Metal, no Python, offline.

Both demos support terminal interaction and an OpenAI-compatible HTTP server.