跪拜 Guibai
← All articles
Backend · Artificial Intelligence · Programmer

Qwen 3.8-27B Lands: A Local 4-bit Model That Matches Opus 4.6 on Coding and Agent Benchmarks

By cxuanAI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A 27B model that runs on a single consumer GPU and competes with frontier closed-source models on coding and agent evals changes the calculus for local dev tooling. Offline, low-latency agent loops with vision and tool-calling become practical without a cloud bill.

Summary

Alibaba's Qwen 3.8 ships two open-weight variants: a 27B dense model with native vision and a 2.4T MoE. The 27B version is the one that fits on consumer hardware. A 4-bit quantized GGUF clocks in at 17.9GB, making it runnable on an RTX 4090, RTX 5080, or a 24GB Mac. Unsloth's Dynamic V3.0 quants and NVFP4 weights for Blackwell GPUs push speed and retention further, with NVFP4 hitting 1.5x BF16 throughput at 92–97% top-1 retention.

Benchmark jumps are concentrated in coding, long-horizon agent tasks, and GUI computer use. SWE-bench Pro rose from 53.5 to 61.7, DeepSWE 1.1 from 13.3 to 42.2, and OSWorld-Verified from 63.9 to 84.3. Those numbers put a locally hosted 27B model in the same conversation as Opus 4.6 Max on several key evals, though GPQA Diamond and HLE still trail.

Deployment paths cover the spectrum: Unsloth Desktop and Studio for a one-click GUI, Ollama and LM Studio for quick pulls, llama.cpp for tool-chain integration, and vLLM/SGLang for multi-user serving with continuous batching. The 2.4T MoE remains a data-center play — its smallest 1-bit quant still wants 450GB RAM.

Takeaways
Qwen 3.8-27B is a dense 27B model with native image and video input, 262K context (extendable to ~1M via YaRN), and thinking mode enabled by default.
The 4-bit GGUF is 17.9GB; total memory needed is 17–19GB, making 24GB GPUs and Macs the realistic floor for serious use.
SWE-bench Pro jumped from 53.5 to 61.7, surpassing Opus 4.6 Max's 53.4; DeepSWE 1.1 rose from 13.3 to 42.2.
Computer-use benchmarks saw OSWorld-Verified climb from 63.9 to 84.3 and WebArena-Verified from 48.8 to 64.8.
NVFP4 weights for Blackwell GPUs (RTX 5090, B200) deliver 1.5x BF16 speed with 92–97% top-1 retention but won't run on RTX 4090 or 3090.
Unsloth Desktop and Studio provide a GUI with built-in thinking, web search, code execution, and tool-calling support.
Ollama serves the official model with one command; LM Studio lets users dial in VRAM offload, context, and sampling parameters manually.
The 2.4T MoE's smallest 1-bit quant is 397GB and still needs ~450GB RAM, keeping it firmly in server territory.
Conclusions

Qwen 3.8's benchmark profile suggests the team prioritized real-world agent and coding workflows over knowledge-recall metrics like GPQA, where scores actually dipped slightly.

The 27B dense model getting vision while the 2.4T MoE is text-only inverts the usual pattern where the largest model gets all modalities.

Unsloth's Dynamic quants and NVFP4 weights arriving at launch show quantization tooling is now a first-class part of a model release, not an afterthought.

Calling a 27B local model 'an Opus 4.6' is marketing shorthand, but the SWE-bench and computer-use numbers make it a defensible comparison for the specific workloads developers care about.

Concepts & terms
GGUF
A file format for storing quantized large language models, used primarily with llama.cpp. It bundles model weights and metadata into a single file that can run on consumer hardware.
KV Cache
Key-Value cache that stores attention computations from previous tokens during inference. Longer context windows produce larger KV caches, consuming more memory.
NVFP4
NVIDIA's 4-bit floating-point format supported on Blackwell GPUs. It offers higher precision than integer quantization at the same bit width and is hardware-accelerated on RTX 50-series and B-series cards.
YaRN
Yet another RoPE extensioN — a method for extending a model's context window beyond its training length by adjusting rotary position embeddings.
MTP speculative decoding
Multi-Token Prediction: a technique where the model predicts multiple future tokens in parallel, then verifies them, reducing latency for generation.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗