Model

Qwen3.8-27B

OPEN

Alibaba · Qwen3.8 family · released Aug 5, 2026

The small open-weight sibling of Qwen3.8-Max, and the one that actually shipped under a permissive licence: Apache 2.0 with no revenue gate, against Qwen3.8-Max's separate-agreement requirement past US$50M of model-as-a-service revenue. RELEASE DATE CONFLICT: BenchLM records August 5, 2026 as an exact-day confirmation against the Hugging Face card, while Capital & Compute reports the weights landing August 14 after Alibaba pointed at the week of August 10; the earlier, primary-source-linked date is used here. A genuinely strong small model — it leads Qwen3.7-Plus (397B/17B) on SWE-bench Pro (61.7 vs 57.6), DeepSWE 1.1 (42.2 vs 14.2), CoWorkBench and JobBench, and posts 84.3 on OSWorld-Verified — while sitting behind it on GPQA Diamond and HLE. Vision rows use the without-code-interpreter setting where Qwen reports both. No first-party per-token API rate is published for this checkpoint, so no pricing row is filed.

ReasoningCodingVisionFunction callingTool useAgentic

2506.1

Elo · rank #30

Parameters

27B

Active params

Undisclosed

Context

262K tokens

Architecture

27B dense causal LM with a vision encoder; 64 layers in a repeating 16 x (3 x (Gated DeltaNet → FFN) → 1 x (Gated Attention → FFN)) hybrid linear/full-attention stack, multi-token prediction, 262,144-token native context extensible to 1M

License

Apache 2.0

Languages

—

API price (in/out)

No hosted API

Modalities

text · vision

Benchmark results

Bar shows position within the tracked field; marker = field best

Agents' Last ExamAgents42.9%#10

best: GPT-6 Astra · 59.3%

AndroidWorldAgents81.9%#2

best: Qwen3.8-Flash-Next · 84.5%

CharXivVision83.7%#9

best: Qwen3.8-Flash-Next · 90.6%

DeepSWECoding42.2%#14

best: Muse Spark 1.3 · 75.4%

GPQA DiamondReasoning89.2%#35

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning30.8%#50

best: Claude Opus 5 · 64.7%

IFBenchReasoning79.5%#10

best: MiniMax M3 · 83.0%

JobBenchAgents33.4%#7

best: Claude Opus 5 · 65.7%

LiveCodeBenchCoding90.3%#6

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MathVisionVision90.0%#4

best: Seed 2.1 Pro · 92.6%

NL2RepoCoding42.3%#4

best: GLM-5.3 · 58.0%

OSWorld-VerifiedAgents84.3%#2

best: Claude Fable 5 · 85.0%

SWE-bench MultimodalCoding38.6%#4

best: Claude Opus 5 · 59.4%

SWE-bench ProCoding61.7%#14

best: Claude Fable 5.1 · 81.2%

Terminal-Bench 2.0Coding73.0%#23

best: Gemini 3.8 Flash · 89.4%

WebArena-VerifiedAgents64.8%#1

best: this model · 64.8%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Fine-tune it

Permissive

QLoRA19.0 GB1× RTX 3090 24GB

LoRA58.2 GB1× A100 80GB

Full fine-tune434.0 GB4× H200 141GB

QLoRA SFT on ~10k samples ≈ $13.91 (1× RTX 3090 24GB)

Plan a fine-tune →

Qwen3.8 family

Elo progression across releases

Qwen3.8 MaxAug 20262863.4 Qwen3.8-27BAug 20262506.1 Qwen3.8-Flash-NextAug 20262659.1 Qwen3.8-Max-0902Sep 20263092.0

Resources

HFQwen/Qwen3.8-27B

API price weights · each benchmark row carries its own source badge (see methodology)