📰
Ornith 1.5 Adds Three Open Models
Ornith AI has released three new Ornith 1.5 model variants: 9B, 35B-A3B, and 397B. Hugging Face listings are available for both the original model formats and GGUF versions.
Search
10 results on this page
⚙️
DFlash2 Hits 138 TPS on RTX 3090
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.
Ornith-1.5 Launches Three Open Model Sizes
Ornith-1.5 launches open models in 397B, 35B, and 9B parameter sizes. The models feature self-improving task and scaffold generation and report strong performance on coding and reasoning benchmarks.
Unsloth Releases Higher-Accuracy Qwen3.8 GGUFs
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.
🤖
Qwen KV Precision Shows Real Quality Gaps
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.
🤖
Qwen3.8-27B Trades Knowledge for Tool Use
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.
Qwen3.8-27B Shows Remarkable Local Agency
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.
A Smaller Qwen3.8 Built by Pruning Layers
A community developer created Qwen3.8-23B-Mini-Me by strategically removing layers from Qwen3.8-27B, reducing the model to approximately 22.7B parameters without severe reasoning degradation. The model is reported to work well for coding, agentic tasks, and multi-turn chats, but it has not yet been benchmarked and struggles more with edge cases and underspecified prompts.
🔬
Gradient Accumulation Can Slow Training
A TraceML benchmark using Qwen3-1.7B with LoRA found that equivalent effective batch sizes can have substantially different training times depending on physical batch size and accumulation steps. Across 100 optimizer updates, 4×1 was 17% faster than 1×4 on a T4 and 41% faster on an L4.
Why Qwen3.8 Traces Aren’t True Reasoning
The post argues that Qwen3.8 intermediate tokens should not be interpreted as human-like reasoning. Citing research, it highlights that trace validity often does not correlate with answer correctness, and that models trained on corrupted or irrelevant traces can perform as well as or better than models trained on valid traces.