Qwen3 27B
A 27-billion-parameter dense large language model in Alibaba's Qwen3 series. It natively supports an extended context length of 262,144 tokens, multimodal input combining text and images, and Multi-Token Prediction (MTP) for inference acceleration.
Core Facts
Qwen3 27B has thinking mode enabled by default, supports disabling thinking, and controls reasoning intensity via a Reasoning Effort parameter.
Qwen3 27B's score on OS World Verified jumped from 63.9 to 84.3 (self-reported official data).
The 4-bit quantized weights of Qwen3 27B are only approximately 16.1 GB, allowing it to run on a MacBook Pro/Mac Studio with 32 GB of memory or a consumer-grade GPU with approximately 17 GB of VRAM.
Timeline (last 90 days)
In a truck seal code tracking task, the summarization baseline hallucinated 11 out of 12 seal codes in the first summary and fabricated 14 out of 18 codes in the third summary; whereas Py CLM correctly reproduced all 18 codes in the same task.
On the BrowseComp-Plus benchmark, CLM paired with Qwen3 27B achieves 59.4%, approximately 11% higher than the best summarization solution, with about one-fifth less computation.
In long-context testing, normal mode ultimately completed up to 114,000 tokens, MTP2 completed up to 81,000 tokens, and failed to complete at 98,000 tokens.
Under conditions of 2944 input tokens and 1024 output tokens, normal mode reaches 29.9 tokens per second, which improves to 48.4 tokens/s after enabling MTP2.
After integrating local Qwen3 into the DeepSeek Harness framework, thinking mode runs at approximately 30 tokens per second, while disabling thinking mode peaks at close to 40 tokens/s for single tasks.
For the same tank mini-game task, local Qwen3 took 4 minutes and 17 seconds, while cloud-based DeepSeek VS Pro took 5 minutes and 54 seconds.
The UP host recommends keeping daily usage around 64,000 tokens, noting that the official spec of 260,000 tokens does not mean a single 24GB GPU can comfortably handle the full context.
After enabling MTP2, total generation time drops from 37.1 seconds to 24.7 seconds, with sustained generation speed increasing by approximately 60%.
This deployment solution can achieve 400+ tokens/s in batch throughput (manual concurrency testing).
Official benchmark data shows that the new version of Qwen3 27B agent coding improved from 13.3 to 42.2, software engineering tasks improved from 49.3 to 79, and frontier agent tasks improved from 10.6 to 20.
40 more timeline events
All Facts (20)
Qwen3 27B has thinking mode enabled by default, supports disabling thinking, and controls reasoning intensity via a Reasoning Effort parameter.
70%VerifiedQwen3 27B's score on OS World Verified jumped from 63.9 to 84.3 (self-reported official data).
70%VerifiedThe 4-bit quantized weights of Qwen3 27B are only approximately 16.1 GB, allowing it to run on a MacBook Pro/Mac Studio with 32 GB of memory or a consumer-grade GPU with approximately 17 GB of VRAM.
65%VerifiedQwen3 27B scored 90.3 on LiveCodeBench V6, slightly surpassing Opus 4.6 Max's score of 88.8.
65%VerifiedQwen3 27B has entered the Ollama official library and can be retrieved with ollama pull qwen3:27b; the default tag is approximately 18 GB, supports a 256k context, and retains text and image capabilities.
65%UnverifiedQwen3 27B uses the Apache 2.0 license and is available on Hugging Face and the ModelScope community for free download, deployment, and commercial use.
60%UnverifiedThe Qwen3 27B architecture uses a 64-layer network with a hidden dimension of 5120, employs a hybrid architecture of Gated DeltaNet and Gated Attention, and consists of units of 16 groups of three linear attention layers plus one full attention layer.
60%UnverifiedQwen3 27B combines three linear attention layers with one full attention layer in a recurrent structure at a ratio of 3:1.
60%UnverifiedDeploying Qwen3 27B with 16 GB of VRAM limits the choice to the Q3_K_XL GGUF quantization version, which is the highest precision that 16 GB of VRAM can accommodate.
60%UnverifiedQwen3 27B achieved a 100% pass rate in the 256K long-context memory retrieval test (needle in a haystack), with all 15 runs being accurate.
60%UnverifiedIn a truck seal code tracking task, the summarization baseline hallucinated 11 out of 12 seal codes in the first summary and fabricated 14 out of 18 codes in the third summary; whereas Py CLM correctly reproduced all 18 codes in the same task.
50%UnverifiedOn the BrowseComp-Plus benchmark, CLM paired with Qwen3 27B achieves 59.4%, approximately 11% higher than the best summarization solution, with about one-fifth less computation.
50%UnverifiedAfter enabling MTP2, total generation time drops from 37.1 seconds to 24.7 seconds, with sustained generation speed increasing by approximately 60%.
50%UnverifiedUnder conditions of 2944 input tokens and 1024 output tokens, normal mode reaches 29.9 tokens per second, which improves to 48.4 tokens/s after enabling MTP2.
50%UnverifiedIn long-context testing, normal mode ultimately completed up to 114,000 tokens, MTP2 completed up to 81,000 tokens, and failed to complete at 98,000 tokens.
50%UnverifiedThe UP host recommends keeping daily usage around 64,000 tokens, noting that the official spec of 260,000 tokens does not mean a single 24GB GPU can comfortably handle the full context.
50%UnverifiedAfter integrating local Qwen3 into the DeepSeek Harness framework, thinking mode runs at approximately 30 tokens per second, while disabling thinking mode peaks at close to 40 tokens/s for single tasks.
50%UnverifiedThis deployment solution can achieve 400+ tokens/s in batch throughput (manual concurrency testing).
50%UnverifiedIn bug-finding tests, the Q6 quantized version of the previous-generation Qwen3 27B found one more bug than the new Q5 version.
50%UnverifiedOfficial benchmark data shows that the new version of Qwen3 27B agent coding improved from 13.3 to 42.2, software engineering tasks improved from 49.3 to 79, and frontier agent tasks improved from 10.6 to 20.
50%