10 results on this page
A hyper-optimized Qwen3.8-27B inference stack using DFlash2 reaches about 138 tokens per second on a power-limited RTX 3090, while 64-request throughput reaches 942 TPS. Prefix caching also reduces long-chat follow-up latency from roughly 23 seconds to under 1.4 seconds in reported tests.
AntLing has released six official Ling-3.0 base checkpoints spanning two model sizes and three training stages. The MIT-licensed, ungated repositories are intended for continued pretraining, fine-tuning, and research rather than direct chat use.
Tencent's Hunyuan Hy4 has reportedly appeared in the model selection list of the Yuanbao app under an expert-level label and with tool-use capabilities. It is positioned above Hy3 and alongside DeepSeek, following Tencent's recent statement that a larger-parameter Hy4 would launch soon with improved performance and multimodal abilities.
An independent builder pretrained a 1.02-billion-parameter Kimi-K3 replica on 5 billion decontaminated tokens for roughly $250. The model achieved 33.4% on HellaSwag, exceeding the reported 28% score of GPT-2 124M despite receiving no instruction tuning.
Unsloth released new Qwen3.8-27B GGUF quantizations claiming 10% higher accuracy at the same file size. The release also includes 1-bit quantizations that retain 77% accuracy and can run with 8GB of RAM.
A builder reports running DeepSeek V4 Flash-0731 at roughly 130–150 tokens per second using 16 RTX 5060 Ti 16GB GPUs connected through two PLX PEX88096 switch islands. The configuration supports up to 500,000-token context with tensor parallelism 8 and up to 1 million tokens with tensor parallelism 4.
Aurora-80K is a tiny language model with exactly 80,000 parameters and a factorized 4,096-token vocabulary. Its reported results include 3.2902 Wikitext-2 BPB, 52.31% BLiMP accuracy, and 26.05% on ARC-Easy.
A community test on an AMD R9700 with ROCm reports noticeable quality and long-context retention differences between FP16 and q8_0 KV cache for Qwen3.8-27B. FP16 reportedly produces more careful structured output and maintains performance beyond 120k tokens, challenging the assumption that both formats are equivalent.
A Reddit user reports that Qwen3.8-27B performs substantially worse than Qwen3.6 on obscure trivia and offline knowledge tests. The model may still be strong for coding and tool-assisted workflows, but users relying on its internal weights for broad factual recall should be cautious.
A Reddit user reports that Qwen3.8-27B autonomously retrieved a university class schedule through 80 tool calls using only credentials and a university name. In another test, it downloaded and analyzed a social-media video, installed Whisper for transcription, and enhanced video frames without human intervention.