Qwen releases Qwen-Drive-1.0-4B, a driving-finetuned open-weight model based on Qwen 3.5 4B

Releases·Model Releasefeatured

Qwen releases Qwen3.8-Flash-Next, previewing Qwen 4 architecture

Qwen released Qwen3.8-Flash-Next, an open-weight hybrid model that previews the Qwen 4 architecture. The 125B MoE model reportedly outperforms DeepSeek V4 Flash at half the parameters and is stronger than Opus 4.6. Its use of n-gram tables could enable running 1T+ parameter models on a single server with modest GPUs and system RAM.

Qwen released Qwen3.8-Flash-Next, an open-weight hybrid model that previews the Qwen 4 architecture. The 125B MoE model reportedly outperforms DeepSeek V4 Flash at half the parameters and is stronger than Opus 4.6. Its use of n-gram tables could enable running 1T+ parameter models on a single server with modest GPUs and system RAM.

Releases·Model Releasefeatured

Prism-ML releases Bonsai Qwen 3.6 27B ternary model for local inference

Prism-ML has released Bonsai Qwen 3.6 27B, a ternary-precision version of the Qwen 3.6 27B model that uses 10x less memory while retaining 95% of the original FP16 benchmark performance. The model fits on 12GB GPUs and is available on Hugging Face, with community members reporting successful fine-tuning using a provided GitHub repository.

Prism-ML has released Bonsai Qwen 3.6 27B, a ternary-precision version of the Qwen 3.6 27B model that uses 10x less memory while retaining 95% of the original FP16 benchmark performance. The model fits on 12GB GPUs and is available on Hugging Face, with community members reporting successful fine-tuning using a provided GitHub repository.

Releases·Model Release

Developer releases QwiVer3.6-35B-A3B, a post-trained Qwen3.6-35B-A3B for local coding

A developer released QwiVer3.6-35B-A3B, a BlackRiver post-trained version of Qwen3.6-35B-A3B, for local coding and agentic workflows. The model retains the original's MoE architecture with ~35B total and ~3B active parameters. The author reports consistently preferring QwiVer over the upstream model after extensive personal use.

A developer released QwiVer3.6-35B-A3B, a BlackRiver post-trained version of Qwen3.6-35B-A3B, for local coding and agentic workflows. The model retains the original's MoE architecture with ~35B total and ~3B active parameters. The author reports consistently preferring QwiVer over the upstream model after extensive personal use.

A new inference engine, Turing Engine, claims to serve 70B-parameter models like LLaMA-3.1-70B and Qwen-2.5-72B on a single 24GB GPU, achieving 3,064 tokens per second with 75% KV cache compression. It supports Unsloth checkpoints, potentially enabling high-throughput inference on consumer hardware.

A model card on Hugging Face for Qwen/Qwen3.8–2.4T-A95B lists a 2.4-trillion-parameter count, which appears to be a typo. The card was fetched on August 16, 2026, and the discrepancy has drawn attention.