Qwen3.8 2.4T-A95B on Run BiOS

Qwen3.8-2.4T-A95B is the open-weight variant of Qwen3.8-Max — the first time a Qwen-Max-class model has been brought to open release. It is a sparse mixture-of-experts causal language model with 2.4 trillion total parameters and 95 billion activated per token. Built on the architectural foundation of Qwen3.5, Qwen3.8 delivers substantial gains across coding, professional work, research, and long-horizon agentic tasks; beyond answering harder questions, it is designed to carry complex, multi-step tasks through to completion with greater reliability.

Qwen highlights four improvements for this generation:

• Core capabilities — comprehensive improvements across coding, professional work, research, and long-horizon agentic tasks.
• Agent execution — stronger autonomous planning and better handling of environment feedback, leading to more reliable end-to-end task completion.
• Downstream compatibility — broader support for popular harnesses and development tools.
• Flexible thinking control — reasoning depth tuned with reasoning_effort, and reasoning context from historical messages retained via preserve_thinking.

Architecture

Published model overview and configuration:

• Scale: 2.4T total parameters with 95B activated per token; BF16 weights.
• Context: 262,144 tokens natively, extensible up to 1,010,000 tokens.
• Layers: 92 hidden layers with a hidden dimension of 8192, arranged as 23 × (3 × (Gated DeltaNet → MoE) → 1 × (Gated Attention → MoE)) — a hybrid layout that interleaves linear attention (Gated DeltaNet) with full gated attention every fourth block.
• Gated DeltaNet: 128 linear attention heads for V and 16 for QK, head dimension 128.
• Gated Attention: 64 query heads and 4 KV heads, head dimension 256, rotary position embedding dimension 64.
• Mixture of experts: 512 experts with 10 routed plus 1 shared expert active per token, expert intermediate dimension 2048.
• Vocabulary: 248,320 (padded), untied input/output embeddings.
• Speculative decoding: a multi-token-prediction (MTP) head trained with multiple steps.

Thinking and Reasoning Effort

Per the official model card:

• Thinking is always on. Qwen3.8-2.4T-A95B requires thinking mode for all interactions — it cannot be disabled. Every response automatically begins with reasoning before the final output.
• Text only. Multimodal inputs are not supported on the open-weight model.
• Reasoning effort ladder. reasoning_effort adjusts reasoning depth and cost: xhigh (the default) for complex tasks demanding thorough analysis, medium for balancing accuracy and speed, and low for efficient reasoning optimizing for speed and cost.
• preserve_thinking is enabled by default for all workloads, retaining reasoning context from historical messages.
• Recommended sampling parameters: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0.

Benchmarks

Benchmark table as published in the official model card. Scores are reported for Qwen3.8-Max, which Qwen describes as the official hosted version based on Qwen3.8-2.4T-A95B.

Coding Agent

Benchmark	Claude Opus 4.8	Claude Fable 5	GPT-5.6 Sol (max)	Qwen3.7-Max	Qwen3.8-Max
Terminal Bench 2.1	84.6	84.6	88.8	74.5	86.6
SWE-bench Pro	69.2	80.0	64.6	60.6	67.7
DeepSWE 1.1	59.0	70.0	73.0	21.6	56.6
NL2Repo-Bench	69.4	-	-	47.2	55.9
FrontierSWE	70.0	88.8	-	40.7	73.5
MLS-Bench-Lite	42.8	49.9	46.2	31.7	41.0
PaperBench	80.3	88.8	90.5	64.8	93.0
AndroidBench	69.8	84.5	74.0	56.5	75.1

General Agent

Benchmark	Claude Opus 4.8	Claude Fable 5	GPT-5.6 Sol (max)	Qwen3.7-Max	Qwen3.8-Max
CoWorkBench	72.3	75.9	71.5	64.6	74.8
WorkSpaceBench	66.8	68.7	65.6	61.4	67.7
JobBench	48.4	57.4	45.4	31.3	53.4
SkillsBench	65.1	70.9	73.5	61.2	70.2
Agents' Last Exam (Pass / Score)	27.0 / 45.1	-	30.6 / 53.6	11.8 / 31.1	27.0 / 52.4
Automation-Bench (Pass@1)	27.2	29.1	29.7	14.2	27.3
Toolathlon Verified (Pass@1)	76.2	77.9	74.9	49.7	72.5
WideSearch	72.9	81.2	-	75.2	81.9
HLE w/ tools	57.9	64.5	58.0	53.5	56.2

General Capabilities

Benchmark	Claude Opus 4.8	Claude Fable 5	GPT-5.6 Sol (max)	Qwen3.7-Max	Qwen3.8-Max
GPQA Diamond	92.0	92.6	94.1	92.4	92.6
HLE	45.7	53.3	47.2	41.4	43.6
IFBench	62.2	63.5	72.7	79.1	82.8
$OneMillion-Bench (expert score)	41.8	55.9	53.8	44.4	52.5
HealthBench	52.4	-	55.3	54.5	60.2
PLawBench	69.6	70.2	72.3	58.9	73.2
PRBench-Legal	52.7	57.6	57.6	48.5	57.6
PRBench-Finance	51.9	55.8	55.5	46.8	58.3
MRCR v2 256K (8-needle)	83.2	-	93.8	86.7	92.9
LongBench v2	69.1	-	67.1	65.3	66.3

Selected evaluation notes as published: Terminal Bench 2.1 was evaluated with Claude Code (avg@10) using a 5-hour timeout and max_tokens=131,072, with other models' best published scores across harnesses; SWE-bench Pro with the Claude Code harness, temp=1.0, top_p=0.95, and a 256K context window; DeepSWE 1.1 with the Claude Code and mini-SWE-agent harnesses (highest score reported); FrontierSWE with the Claude Code harness against official MEAN@5 leaderboard results as of August 3, 2026; empty cells mean scores were not yet available or not applicable. Claude Fable 5 results may involve fallbacks.

License

Qwen3.8-2.4T-A95B is released under the Qwen3.8-Max License: a permissive grant (use, modify, distribute, sell, deploy, host, fine-tune, and create derivative works) with two conditions — attribution/prominent model-name display for products exceeding 100M monthly active users or US$20M monthly revenue, and a separate license requirement for Model-as-a-Service or AI Work Assistant businesses whose aggregate revenue exceeds US$50M over any consecutive twelve months.

On Run BiOS

On the Run BiOS serverless catalog this model is served as qwen3.8-2.4t-a95b with a 262,144-token advertised context window (the model's native length; individual routes serve up to 1M tokens), always-on thinking (per the vendor, reasoning cannot be disabled), adjustable reasoning effort exposed as a low / medium / max ladder (max runs at the vendor's xhigh level, its own default), function calling, structured outputs, and prompt caching for repeated prefixes.