~ / guides / qwen3.8-27b local

Qwen3.8-27B: a 28B coder on one graphics card

Alibaba published Qwen3.8-27B on 13 August 2026 under Apache 2.0. At 4-bit it fits in 14–17 GB, which means a single RTX 4090 runs it. This page covers what the hardware actually has to be, how to start it, and how far behind the frontier it really is — measured, not implied.

28BParameters, dense

14–17 GBVRAM at 4-bit

262KNative context

Apache 2.0Licence

42.2DeepSWE pass@1

Parameters and the card you need

Weights only. Read the caveat under the table before you buy anything.

Precision	Weights in memory	Card class	Cards that work
4-bit GGUF / AWQ	14–17 GB	24 GB	RTX 4090, RTX 5090
FP8	~28 GB	48 GB	L40S, RTX Pro 6000, RTX 5090
BF16 (full)	~56 GB	80 GB	H100, H200, RTX Pro 6000 96GB
4-bit on Apple Silicon	~24 GB unified memory	—	M-series with 24 GB or more

The table is weights only. KV cache sits on top of every figure and grows with context length and with how many requests you run at once. On a 262K-context model that is not a rounding error — a long session or a couple of concurrent agents can push real usage toward double the number above. If you are sizing a 24 GB card at 4-bit, plan on short-to-mid context, not the full 262K.

Starting it

Three routes, easiest first.

Ollama

ollama pull qwen3.8:27b
ollama run qwen3.8:27b

llama.cpp, if you want control over context and quantisation

llama-server -m qwen3.8-27b-Q4_K_M.gguf -c 8192 --port 8080

Q4_K_M has been the community default for this size class. Lower quants free memory but the quality loss on long agent runs shows up sooner than on short chat turns.

vLLM, for serving several users

vllm serve Qwen/Qwen3.8-27B

Official weights: Qwen/Qwen3.8-27B on Hugging Face, Apache 2.0 — commercial use, modification and redistribution all permitted.
The model ships a vision encoder, so it takes images as well as text.
It speaks the OpenAI-compatible Chat Completions API and supports tool calling and JSON output, so most agent harnesses point at it without changes.

Coding ability against the frontier

DeepSWE pass@1 — share of long-horizon software-engineering tasks finished on the first attempt. Same benchmark for every bar.

DeepSWE pass@1

Higher is better · one benchmark, so these numbers are directly comparable

Claude Opus 573.6
Kimi K368.5
GPT-5.567.0
DeepSeek V4 Pro 081362.7†
Qwen3.8-27B42.2

† DeepSeek V4 Pro 0813's 62.7 is DeepSeek's own published figure. Independent trackers recorded no third-party evaluations at publication, and the agentic harness behind it has not been released, so the result cannot be reproduced yet. Opus 5, Kimi K3 and GPT-5.5 are measured entries on the 7 August DeepSWE table. Qwen3.8-27B's 42.2 comes from Alibaba's model card.

Model	DeepSWE	Gap vs Qwen3.8-27B	Weights	Memory to self-host
Claude Opus 5	73.6	+31.4	closed	API only
Kimi K3	68.5	+26.3	open, bespoke licence	~1.4 TB, 64+ accelerators
GPT-5.5	67.0	+24.8	closed	API only
DeepSeek V4 Pro 0813	62.7†	+20.5	MIT	~430 GB
Qwen3.8-27B	42.2	—	Apache 2.0	14–17 GB, one 24 GB card

Read the two columns together. Qwen3.8-27B finishes roughly 42 tasks in 100 where Opus 5 finishes about 74. That gap is real and this page is not going to talk you out of it. What the score column does not show is that the four models above it are either closed weights you can only rent, or open weights whose hardware bill starts at a 430 GB server and ends at a 64-accelerator cluster. Qwen3.8-27B is the only row on this table that runs on a card you can buy for a workstation.

What changed since Qwen3.6-27B

Same size class, one generation apart.

Qwen3.6-27B to Qwen3.8-27B

Alibaba's published comparison across four benchmarks

DeepSWE42.2was 13.3
Terminal Bench73.0was 63.4
SWE-bench Pro61.7+8.2
OSWorld84.3was 63.9

DeepSWE more than tripled between the two generations, from 13.3 to 42.2. These four benchmarks measure different things and their numbers are not comparable to each other — only each figure against its own predecessor.

Where the evidence comes from

So you can weigh each number rather than take the page's word for it.

Measured on the independent table: Claude Opus 5 (73.6), Kimi K3 (68.5), GPT-5.5 (67.0) — DeepSWE refresh of 7 August 2026.
Vendor-published, not yet reproduced: DeepSeek V4 Pro 0813 (62.7) and Qwen3.8-27B (42.2). Both come from the model owner. Treat them as claims until a third party runs them.
Hardware figures are weights-only memory at each precision, from the published requirements. Throughput on your own machine depends on quantisation, context length and whether speculative decoding is on, so this page gives no tokens-per-second number it cannot stand behind.

See the full local hardware table Full DeepSWE leaderboard