Qwen 3 Coder Next

Qwen 3 Coder Next is a coding-specialized Open Weights MoE model by Alibaba with 80 billion total and 3 billion active parameters. Q4 quantization significantly reduces memory requirements for local inference; the context window spans 262,000 tokens. Deployable locally on Workstation hardware under the Apache 2.0 license, optimized for coding agents and large codebases.

Alibaba Version 3 Coder Commercial use permitted MoE 80 B (3 B active) 262 K Context 05/2025 locally tested

Open Weights
Workstation
llama.cpp
Text
Instruction-Tuned
Agentic Orchestrator
Interactive

Tool-use details

Asset performance, reliability, and runtime profile

Name	P1	P2	Combined
EU License Research	100	60	76
HTTP Fetch & Extract	80	35	57.5
Tool Failure Handling (404)	80	80	80
Web Search & Tool Selection	100	15	61.8
URL Construction & Fetch	80	60	70
Multilingual Search & Synthesis	100	60	80

Name

Combined

EU License Research

100

HTTP Fetch & Extract

57.5

Tool Failure Handling (404)

Web Search & Tool Selection

100

61.8

URL Construction & Fetch

Multilingual Search & Synthesis

100

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

Reliability

Tool Call Valid: Yes

Retry: Not required

Hallucination: 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: 1.58; First Request

MCP: 0.84; Protocol Latency

Synthesis: 12.74; Response Generation

Total: 90.98; Sum of All Phases

Token: 9962; 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.

Deployment Verdict

Created on: 14.06.2026, 16:09:25

Conditional deploy, because tool execution is robust, tool calls remain valid, and the overall impression is good — but synthesis quality with detected hallucinations is not stable enough for unsupervised high-trust pipelines.

Tool Execution Profile

This model can generally be trusted in an MCP tool chain. It selects tools not merely by rote, but demonstrates genuine selection intelligence: in the Web Search & Tool Selection test, it correctly recognizes without an explicit hint that a search is required before a direct fetch. This points to agentic behavior in open retrieval flows. In the URL Construction test — which requires deriving the correct target URL from internal knowledge and then retrieving it via fetch — it performs adequately, but not deterministically enough for pipelines with strict URL precision requirements. The P1 values thus reveal a clear profile: high protocol adherence, good tool selection, a slight weakness in the exact pre-retrieval step. Also notable: no retry was needed. That is a comprehension signal, not merely a format hit.

Synthesis Fidelity

How well does it condense tool results? Only with limited reliability. Execution is strong, but downstream condensation remains the bottleneck. Precision drops noticeably in particular during the HTTP Fetch & Extract test, which requires pulling structured facts such as dates and proper nouns from real page content. In the Web Search & Tool Selection test as well, tool selection was correct, but synthesis of the retrieved material was weak. For productive tool pipelines, this means: the model often finds the right path, but does not formulate the result with consistent precision.

Does it stay within tool output or fall back on training data? In the Honeypot EU License Research — which tests whether current license restrictions are genuinely drawn from web sources — it stays within the tool result space. That is the most important trust signal here. At the same time, a globally detected hallucination finding exists in the run. This is not merely a quality deficiency but a security risk: when a model presents fabricated facts as tool output, it undermines the reliability of the entire infrastructure.

Error Resilience

The model responds to tool errors in a production-appropriate manner. In the 404 test, which checks for transparent behavior on failed retrieval, it communicates the error rather than fabricating page content. This is precisely the behavior that is acceptable for productive agents. The finding is clearly positive.

Sovereignty Profile

Locally operable and practically deployable. With a Combined Score of 70.88, it sits 1.37 points above the fleet average of 67.84. For a local Q4-GGUF variant, that is a strong sovereignty value — particularly because tool execution does not visibly collapse under quantization.

Conclusion & Recommendation

Suitable for local coding and agent pipelines where tool navigation, web research, error transparency, and MCP conformance matter more than perfect final condensation. Not suitable for compliance, policy, or executive summary workflows without downstream validation. Recommended as a worker model with guardrails: tool-first, keep citations or raw results visible, and either verify final synthesis or hand it off to a stronger condensation model.

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.