GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

Qwen3.6-Plus Claims 78.8% SWE-Bench at Half the Scale of Its Rivals

Alibaba’s Tongyi Lab launched Qwen3.6-Plus on July 5, the closed-API flagship of the Qwen3.6 family, with a specific claim: the first model in its class — smaller than Kimi K2.7 and GLM-5.2 — to claim top-tier agentic programming performance across coding, debugging, and terminal automation.

Vendor-reported benchmarks: 78.8% SWE-Bench Verified, 61.6% Terminal-Bench 2.0, 1 million token context window. Those numbers are self-reported and harness-dependent. SWE-Bench Verified scores are extremely sensitive to agent scaffolding — a 3-4 point spread between labs on the same underlying model capability is noise, not signal.

The credibility check: Qwen3.6-27B, the open-weight dense model from the same family, lands at 77.2% on SWE-Bench Verified, a figure that has been corroborated by independent analysis. The Plus model’s 78.8% claim is plausible if it represents an incrementally tuned version of that same architecture with proprietary post-training.

What Changed Architecturally

The Qwen3.6 series uses a hybrid mixture-of-experts backbone. The Plus model introduces a preserve_thinking API parameter that retains chain-of-thought across turns — addressed specifically at multi-step agent tasks where planning context degrades across session boundaries. That is the meaningful architectural addition. Without it, long-horizon agent tasks that require planning state to persist across dozens of tool calls degrade as context fills.

On multimodal tasks, the model closes a visual-to-execution loop: it can take a UI screenshot, identify an interface element, and generate code to interact with it. The benchmark Tongyi cites is a scratch-card tally task — trivial but representative of the class of problems where vision feeds agentic execution.

Compatibility Play

Qwen3.6-Plus ships with Anthropic protocol compatibility. Developers using Claude Code can route calls to the Qwen API without changing their tooling setup. That is not a capability claim — it is a distribution play. Claude Code’s CLAUDE.md convention and toolchain are the de facto standard for production coding agents in 2026. Matching the protocol removes the switching cost for price-sensitive developers.

Qwen Code, Alibaba’s own terminal agent fine-tuned for the Qwen family, is the recommended native interface.

Benchmark Position

For comparison across the Qwen3.6 line:

ModelSWE-Bench VerifiedTerminal-Bench 2.0Params
Qwen3.6-Plus78.8% (vendor-reported)61.6%Undisclosed
Qwen3.6-27B77.2%59.3%27B dense
Qwen3.6-35B-A3B73.4%—35B MoE (~3B active)

The Plus model leads its open-weight siblings by 1.6 points on SWE-Bench but only 2.3 points on Terminal-Bench 2.0. That is a thin margin. The open-weight Qwen3.6-27B, which runs on a single consumer GPU at quantized precision and costs nothing per API call, is within measurement noise.

The frontier ceiling for context: Gemini 3.5 Flash leads Terminal-Bench 2.1 at 76.2%, GPT-5.5 leads at 84.7% (third-party agent), and Fable 5 sits at approximately 84.3%. Qwen3.6-Plus at 61.6% on the older Terminal-Bench 2.0 harness is not a frontier coding claim — it positions the model in the second tier, competitive with mid-2025 frontier models and notably ahead of open-weight alternatives at similar scale.