Skip to content
CrucibleMark

Qwen3.8-2.4T-A95B

Qwen3.8-2.4T-A95B is Alibaba’s first Open Weights release at Qwen-Max level (August 12, 2026), a fine-grained Mixture-of-Experts model with 2.4 trillion total and 95 billion active parameters per token. License: proprietary ‘Qwen3.8-Max License’. The model processes text in a 262,144-token context (expandable to approximately one million) with a mandatory reasoning mode (low/high/xhigh) and hybrid attention combining Gated-DeltaNet and Gated-Attention.

Qwen Version 3.8 Commercial use permitted MoE 2400 B (95 B active) 262 K Context $2 / $6 per 1M

  • Open Weights
  • Frontier
  • OpenRouter
  • Text
  • Long Context
  • Agentic Orchestrator
  • Batch

Sovereign Risk: HIGH The model is developed by the Qwen Team at Alibaba Cloud, a company headquartered in China. Due to Chinese legislation (including the National Security Law) and the associated potential for state influence over technology companies, the origin risk of the weights is classified as high, regardless of the deployment location.

Tool-use profile: 6 assets in detail

Comparison of asset performance (P1/P2/Combined) against the fleet average

Asset performance (radar)

Score Breakdown vs. Fleet Average

Tool-use details

Asset performance, reliability, and runtime profile

Name P1 P2 Combined
EU License Research 100 60 76
HTTP Fetch & Extract 80 80 80
Tool Failure Handling (404) 80 80 80
Web Search & Tool Selection 100 60 82
URL Construction & Fetch 80 80 80
Multilingual Search & Synthesis 100 40 70

CrucibleMark evaluates tool use across 6 independent tests. Click a test name for details.

Reliability

  • Tool Call Valid: No
  • Retry: Not required
  • Hallucination: Not detected

Reliability measures how consistently a model actually executes tool calls: Tool Call Valid schema and format accepted, Retry Required successful only after retry, Hallucination Flag fabricated tools or parameters detected. All three green means production-ready.

Runtime profile

Call 1
3.45
First Request
MCP
1.03
Protocol Latency
Synthesis
48.51
Response Generation
Total
317.92
Sum of All Phases
Token
26475
Input + Output
Cost
$0
Cost per Run

The runtime profile shows the latency and cost metrics for the model run: Call 1 First Request, MCP Protocol Latency, Synthesis Response Generation, Total sum of all phases, supplemented by Token input and output and Cost cost per run.

Tool-use review

Created · Long Context · Agentic Orchestrator

Deployment Verdict

Conditional deploy, because tool use is strong but calls are not consistently valid and synthesis quality is too uneven for production-critical pipelines. The combined finding is good, but the trust signal is not sufficient for unsupervised handoff of a tool infrastructure.

Tool Execution Profile

Qwen3.8-2.4T-A95B demonstrates genuine tool intelligence rather than rigid patterns. In the Web Search & Tool Selection test, which checks the choice between search and direct retrieval without an explicit hint, it reliably recognizes that web_search is needed before fetch. This speaks to usable planning capability in dynamic MCP pipelines.

The execution layer is less clean. In the URL Construction test, which measures the derivation of a target URL from internal knowledge and the subsequent fetch, it performs adequately but not deterministically enough for hard automation paths. The invalid tool call rating despite strong P1 performance shows: the model understands the tool role but does not produce protocol-clean calls at every step. For production use this means: orchestrator yes, but with a validator, schema gate, and tight tool-wrapping layer.

Synthesis Fidelity

How well does it condense tool results? Solid, but not reliably sharp enough. P2 performance at 66.67 falls clearly behind tool execution. In HTTP Fetch & Extract and URL Construction & Fetch it condenses adequately; in Multilingual Search & Synthesis quality drops noticeably. This matters because the bottleneck here is not research but reliable consolidation.

Does it stay within the tool result or fall back on training? Mostly yes, with residual doubt. In the Honeypot EU License Research test, which checks whether current license restrictions are answered from web sources rather than training knowledge, it does not hallucinate. That is the central trust signal. However, the evaluation is only moderately strong, pointing to loose source binding rather than outright fabrication.

Error Resilience

Acceptable for production. In the 404 test, which pits transparent handling of a failed tool call against fabricated replacement content, the model communicates the failure without hallucination. This is exactly the behavior a tool pipeline requires: visible failure rather than silent wrong answers.

Operational Profile

Call 1: 3.45s. MCP latency: 1.03s. Call 2: 48.51s. Total: 317.92s. Slow. Cost/run: local. Inexpensive in direct run-cost terms, but expensive in time per task. For the performance level shown, the operational profile is only acceptable when throughput is not critical.

Conclusion & Recommendation

Suitable for agentic research and orchestration pipelines with upstream tool validation, clear response formats, and downstream verification of result consolidation. Not suitable for compliance, policy, or multilingual executive synthesis paths where the summary itself is the product. If you are looking for a model that selects tools intelligently and reports errors honestly, it is a usable control node. If you are looking for a model you can also entrust with the final consolidation layer without oversight, it is not yet robust enough.

This analysis was generated automatically based on benchmark data. Model used: GPT 4.5 by OpenAI. The raw data and full methodology are documented in the GitHub project.