Release of Qwen 3.8-27B Model

The Qwen team has officially confirmed the release of the Qwen 3.8-27B large language model. While initial speculation on community forums like Reddit suggested a potential delay, technical details surfaced via ModelScope, indicating the model’s availability.

According to technical observations shared by users on Reddit, the 3.8 version of the 27B model maintains an identical architecture to its predecessor, Qwen 3.6-27B. This suggests that the performance improvements in the new iteration are primarily the result of enhanced training methodologies rather than structural changes to the model’s framework.

Performance Benchmarks and Intelligence Indices

Early assessments of the model’s capabilities suggest significant competitive positioning. Unverified reports from community discussions indicate that Qwen 3.8-27B ranks second on Artificial Analysis’s intelligence index, trailing only the Kimi K3 model by approximately two points. Notably, the model reportedly holds the top position among open-weight models on the same platform’s agentic index, surpassed only by Opus 5.

However, some users have noted that the model may exhibit a tendency toward “overthinking” certain queries. Tech commentator Simon Willison noted that while the model is highly engaging for local deployment, it can default to excessive reasoning unless specific prompting adjustments are applied to streamline its output.

Hardware Performance and Inference Speeds

Preliminary performance data from independent testing provides insight into the model’s efficiency on high-end consumer hardware. In a series of tests conducted on an NVIDIA RTX 5090, the model demonstrated high throughput across varying context lengths.

According to unverified logs from a single-session Codex run, the median decoding speeds were recorded as follows:

  • Under 50k context: 169.7 tokens per second (tok/s)
  • 50k–100k context: 167.6 tok/s
  • 100k–150k context: 156.9 tok/s
  • 150k–200k context: 149.0 tok/s
  • Over 200k context: 143.7 tok/s

The testing showed that while speed decreases as the context window expands, the degradation is relatively gradual. The maximum observed decode speed reached 222.0 tok/s. These tests were conducted using a specific quantization (NInfer) with the KV cache set to q8, providing a baseline for how the model might perform in local, high-memory environments.