The Latest Top LLMs: The Qwen3.8 Family, Specialized Releases, and Inference Software

A quick, transparent note before the roundup: this post reflects research gathered as of August 2026 and focuses on the Qwen3.8 family, which is the most heavily documented current release. Coverage of other top models (e.g., Fable 5, Kimi K3) is necessarily brief, and live leaderboard sites such as LM Charts did not load in this research session, so some rankings are described at a high level rather than with full tables.

Why Qwen3.8 matters

Qwen (Tongyi Qianwen) is Alibaba Cloud’s family of predominantly open-weights language models. Its latest version, Qwen3.8, is the first in the series to bring a “Qwen-Max-class” model to open release, built on the architectural foundation of Qwen3.5 (the hybrid Gated DeltaNet + Gated Attention design). On independent agentic and coding benchmarks it ranks among the top open-weights models worldwide, and Qwen has become the community’s de-facto base model for fine-tuning and deployment.

Read the official blog: qwen.ai/blog, and the series overview: github.com/QwenLM/Qwen3.8.

The Qwen3.8 family

Qwen3.8-Max — Alibaba’s flagship (announced Aug 3, 2026), reported at 2.4 trillion parameters with 95 billion active, and the largest/most capable Qwen model to date. Its core claim is “second only to Fable 5” on agentic and science-benchmark data (though the engineering margin is smaller). Note: Alibaba also released the open weights of its flagship, Qwen3.8-2.4T-A95B, on Hugging Face and ModelScope. Details are drawn from a Chinese-language model wiki and should be treated as vendor claims pending broader corroboration. See heyaiwiki.
Qwen3.8-2.4T-A95B — The open-weight flagship (2.4T total / 95B active), up to 1 million-token context. Reported third globally on Arena AI’s CodeArena and third on Artificial Analysis’ Agentic Index.
Qwen3.8-27B — The main open-weight, native multimodal dense model (understands images and video). ~27B parameters (~28B with the ~1B vision encoder), 64 layers, hidden size 5,120, and a 248,320-token vocabulary. Released under Apache 2.0, runs on consumer hardware and, with quantization, on a laptop. Native 262K-token context that extends to 1 million tokens. Matches the performance of Qwen3.7-plus (an MoE model ten times its size). Officially launched Aug 14, 2026, and became one of the top 5 most-liked models on Hugging Face within two days. Alibaba framed it as a challenge to Meta for “laptop-ready” models. See the Alibaba Cloud blog: alibabacloud.com/blog/alibaba-unveils-qwen3-8-27b_603463 and CNBC.
Qwen3.8-Flash-Next — A multimodal MoE model that serves as an early preview of the architecture used in Qwen4. See openlm.ai/qwen3.8.

Broader context: other recent top LLMs

Kimi K3 (Moonshot AI, China) — released just before Qwen3.8 and was the top open-weights Chinese LLM as of Aug 12, 2026, per Wikipedia.
Fable 5 — referenced by Alibaba as ranking above Qwen3.8-Max on agentic/science benchmarks (limited independent detail available in this session).

Caveat: full cross-vendor benchmark tables could not be retrieved during research, so treat the above as directional rather than exhaustive.

Specially trained / community releases

These are third-party fine-tunes and adapters built on top of Qwen3.8 (surfaced via the Hugging Face search for qwen3.8). They are not official Alibaba releases:

HauhauCS/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF — Image-Text-to-Text, 2B, updated Aug 17.
ttindall/Qwen3.8-27B-Defiance-Uncensored-MTP-GGUF — Text Generation, 27B, updated ~8 days prior.

The -GGUF suffix means these are quantized, CPU/llama.cpp-friendly formats (4-bit/8-bit, and various quantization levels), which is what lets a 27B model run on local hardware. This is the same reason Qwen3.8-27B can run on a laptop.

Optimized software that serves these LLMs

Qwen-specific: QwenWork (enterprise workflows) and Alibaba Cloud Model Studio (one-stop generative-AI platform); plus Qwen-MM-Plugins, Alibaba’s harness extension library for multimodal agents (image/video, multimodal memory, dynamic-resolution, visual tool use).
General open-weight inference servers (from the Hugging Face inference-provider list): vLLM, SGLang, Ollama, MLX, LM Docker, Unsloth, and LM Studio. These let you run Qwen3.8 (and other open models) locally or self-hosted at scale. See qwen.ai/blog for the official guidance.

Where Qwen stands in the ecosystem