Qwen3.8-27B is emerging as a viable local model for agentic coding, with users reporting multi-hour autonomous sessions, visual feedback loops, and reduced API costs.
An unverified report describes a nearly twenty-hour agentic coding session with Qwen3.8-27B Q6 on an RTX 3090 and an RTX 3060, sustaining 60–63 token/s. If replicated, the data shifts the cloud versus self-hosted boundary. We analyze trade-offs, TCO ...
A user reports nearly twenty hours of agentic coding with Qwen3.8-27B Q6 on an RTX 3090 and an RTX 3060, sustaining 60–63 tokens/s. The case shows how a mid-size quantized LLM can handle prolonged on-prem workloads on consumer hardware, shifting the ...
A test on a 16GB RTX 4060 Ti shows Qwen 3.8 27B at 30-35 tokens/s with Q3_xxs quantization, strong coding and logic, but conversational slips. The Q4 floor is not automatic for local deployments.
On a 32GB RTX 5090, a new GGUF NVFP4 quant for Qwen3.8 27B reaches 6,250 tokens/s in prefill with 2048-token prompts, 50% faster than a Q4_0 of the same memory footprint and 4-7% faster than other NVFP4 quants. It includes a quantized MTP draft head ...
The expiration of a Claude Code Pro subscription pushed a user to a local Qwen3.8-27b LLM on a 5090M GPU with 24GB of VRAM, alongside Pi. A test app for aurora forecasting showed similar timing, a better UI from Pi but better science from Claude Sonn...
Hands-on tests and offline benchmarks suggest Qwen3.8-27B performs worse than Qwen3.6 on factual recall when no external tools are used. For air-gapped deployments relying on model weights alone, the regression is significant.
Unsloth has published new Qwen3.8-27B GGUF files based on Dynamic v3.0. The company reports more than 10% higher accuracy at the same size and a 1-bit quantization retaining 77% accuracy while running on 8GB of RAM. It clarifies the update is not a f...
A bare-metal test with two RTX 3090s, vLLM, and DFlash2 speculative decoding pushes Qwen3.8-27B to 218 tok/s on a single request, with prefill up to 1342 tok/s and a 131k context ceiling. The setup uses INT4 quantization, custom vLLM changes, and pea...
An agentic workload running for over eight hours on a single RTX PRO 6000 with DeepSeek Harness and NInfer handled 966 model calls, 131.2 million input tokens and 853.3 thousand output tokens with zero generation failures. The API price comparison es...
An amateur comparison between Qwen3.6-27B and Qwen3.8-27B on a BASIC ray tracer shows a decisive difference: the ability to observe rendered output and correct code autonomously. With aggressive quantization and local hardware, a closed loop cuts hum...
A hobbyist compared two 27B-parameter LLMs with unsloth UD-Q8_K_XL quantization in an agentic harness: write a recursive ray tracer in BASIC, run it, inspect the image, and iterate. Qwen3.6 needed human input when it couldn't see the mistake; Qwen3.8...
A local model on a Framework Desktop with Q8 GGUF quantization one-shots a Super Mario clone. It is not fast, but smart enough for overnight batches and background jobs. The case raises concrete questions about speed, accuracy, and on-premise deploym...
A commit in the ms-swift framework exposes the string Qwen 3.8 35BA3B, with no announcement or specs. The name suggests a 35-billion-parameter model with a mixture-of-experts architecture, but the source confirms nothing. We analyze what it means for...
