Qwen 3.8 27B
This machine runs a 27.8-billion-parameter dense model. On the Artificial Analysis Intelligence Index v4.3, Qwen 3.8 27B (xhigh) scores 34. Gemini 3.6 Flash is 34. Claude Sonnet 5 (xhigh) is 35. A 4-bit copy of the weights is 17 GB, Apache 2.0, on Hugging Face.
- Intelligence
- 34
- Parameters
- 27.8B
- Q4 on disk
- 17 GB
Intelligence Index
Artificial AnalysisClosed models are APIs with unpublished size. The 27B is a file. Full table below.
The gap
Parameters and intelligence
The local dense model is 27.8 billion parameters. Comparable U.S. mid-tier APIs — Sonnet and Gemini Flash — sit nearby on the scoreboard. The Artificial Analysis Intelligence Index v4.3 is a separate number. It is listed next to each model so size and score can be read together.
Peer gap
1 pt
Claude Sonnet 5 (xhigh) scores 35. Qwen 3.8 27B (xhigh) scores 34 — one Index point apart on v4.3.
Gemini 3.6 Flash
34
Qwen 3.8 27B scores 34. Gemini 3.6 Flash scores 34. Claude Sonnet 5 (high) scores 32.
Same size, one generation
+12
Qwen 3.6 27B scored 22. Same ~17 GB Q4. Twelve Index points to Qwen 3.8 27B at 34.
Parameters, log scale
2021
2023
2024
2025
2026
Intelligence Index
Artificial Analysis v4.3- Qwen 3.6 27B
27B · 17 GB Q4 · Reasoning22 - Qwen 3.8 27B
27.8B · 17 GB Q4 · medium28 - Claude Sonnet 5
closed · API · high32 - Gemini 3.8 Flash
closed · API · low34 - Gemini 3.6 Flash
closed · API · high34 - Qwen 3.8 27B
27.8B · 17 GB Q4 · xhigh34 - Claude Sonnet 5
closed · API · xhigh35 - Claude Sonnet 5
closed · API · max38
Size
The file got smaller. Intelligence did not.
A good local model in 2023 was a 130 GB FP16 checkpoint. The current default is a 17 GB Q4 that scores 34 on the Intelligence Index v4.3 — twelve points above Qwen 3.6 27B (22), at the same disk size.
2023 · Llama 65B FP16
2024 · Llama 3 70B Q4
2026 · Qwen 3.6 27B Q4 · Index 22
2026 · Qwen 3.8 27B Q4 · Index 34
Quantization
Bits traded for VRAM
The same 27.8B weights, stored at different precisions. File sizes from Unsloth GGUFs. Memory bands are Unsloth’s published RAM/VRAM ranges.
Token generation
Decode rates on local hardware
Generation speed for Qwen 3.8 27B at Q4. The model scores 34 on the Intelligence Index v4.3, matching Gemini 3.6 Flash and one point under Claude Sonnet 5 (xhigh) at 35. A 4090 at Q4 is about 49 t/s in llama.cpp. A 5090 with multi-token prediction is reported around 100 t/s.
llama.cpp decode, published 4-bit figure
llama.cpp with multi-token prediction, fresh context
Reported ~100 t/s on a single 5090
Mac-only decode around 14 t/s; ~22 t/s with MTP
Hardware
Weights plus the KV cache
Disk size is not the VRAM bill. The 16 full-attention layers add about 64 KB per token at FP16. The other 48 layers keep a fixed-size DeltaNet state.
Architecture
Why 27 billion is enough here
Dense, not sparse
Every one of the 27.78 billion parameters activates on every token. No expert routing. Predictable latency, full quality at this size.
Gated DeltaNet
Forty-eight of sixty-four layers keep a fixed-size state. Only sixteen layers store a per-token attention cache — about 64 KB per token at FP16, not 256 KB.
Hybrid thinking
Reasoning traces are first-class. Effort is tunable — xhigh, medium, low, or none — so a 24 GB card is not forced to overthink a greeting.
Native vision
Images, diagrams, documents, and hour-scale video in the same weights. No separate encoder to juggle at load time.
Run it
How the weights are run
Alibaba published Qwen 3.8 27B under Apache 2.0 on Hugging Face. Chat on this machine is Open WebUI. This page does not host the model.
Easy
Unsloth
Unsloth publishes Dynamic GGUFs of Qwen 3.8 27B and a desktop app that loads them. Choose a quant that fits the machine. No compile step.
Medium
Ollama
Ollama pulls a GGUF and serves an OpenAI-compatible endpoint. Open WebUI is the chat frontend on this machine.
Expert
llama.cpp
Same GGUF files, more flags: KV cache type, YaRN, multi-token prediction, CPU/GPU split.