Qwen 3.6 vs 3.5: Full Benchmark Table + Should You Upgrade?

Qwen 3.6 35B-A3B beats the previous Qwen 3.5 35B-A3B release on every benchmark in Alibaba’s official model card, with the largest gains in agentic and tool-use tasks. It scores 73.4 on SWE-bench Verified (up from 70.0), 51.5 on Terminal-Bench 2.0 (up from 40.5), and 37.0 on MCPMark (up from 27.0), while keeping the same mixture-of-experts efficiency of roughly 3 billion active parameters out of 35 billion. There is no benchmark regression: reasoning and math improve too.

Benchmark comparison: Qwen 3.6 vs Qwen 3.5

All figures below come from the official Qwen3.6-35B-A3B model card↗ and compare the two 35B-A3B releases head to head. Both models share the same parameter budget, so this is a clean generational comparison rather than a size mismatch.

Benchmark	Qwen 3.6 35B-A3B	Qwen 3.5 35B-A3B	Change
SWE-bench Verified↗ (coding)	73.4	70.0	+3.4
Terminal-Bench 2.0 (agentic)	51.5	40.5	+11.0
MCPMark (tool use)	37.0	27.0	+10.0
LiveCodeBench v6↗ (coding)	80.4	74.6	+5.8
GPQA Diamond (reasoning)	86.0	84.2	+1.8
AIME 2026 (math)	92.7	91.0	+1.7

How to read these numbers

Benchmark deltas only mean something if you know what each test measures. SWE-bench Verified checks whether a model can resolve real, human-validated GitHub issues end to end, so it is the closest proxy for day-to-day software work. Terminal-Bench 2.0 measures agentic competence in a real shell: running commands, reading output, and recovering from failures over many steps. MCPMark scores how reliably a model uses external tools through the Model Context Protocol. LiveCodeBench v6 is a contamination-resistant coding test that refreshes its problems over time. GPQA Diamond is a graduate-level science reasoning set, and AIME 2026 is competition mathematics.

Read together, the pattern is clear. The two reasoning and math tests (GPQA, AIME) were already high on Qwen 3.5 and inch up a point or two, while the agentic and tool-use tests (Terminal-Bench, MCPMark) jump by double digits. Qwen 3.6 is not a small across-the-board polish. It is a targeted upgrade aimed squarely at agent and tool workloads.

Where Qwen 3.6 pulls ahead: agentic and tool use

The headline story is not raw coding, it is agentic work. Terminal-Bench 2.0 climbs 11 points (40.5 to 51.5) and MCPMark climbs 10 points (27.0 to 37.0). In practice that means the model is far better at multi-step tasks: running a command, reading the result, deciding the next action, and recovering when something fails partway through a session. Those are the exact failure modes that make older models frustrating inside agent loops, where a single bad tool call can derail a whole run.

If you are building agents, wiring the model into MCP servers, or running it behind a coding assistant that executes shell commands, this is the upgrade that matters most. A model that scores 11 points higher on Terminal-Bench is one that needs fewer retries and less babysitting per task.

Coding gains

On coding, SWE-bench Verified rises from 70.0 to 73.4, and LiveCodeBench v6 rises from 74.6 to 80.4. The SWE-bench gain is the more meaningful of the two because it reflects real repository work rather than isolated puzzles: setting up the change, editing the right files, and producing a patch that actually passes the project’s tests. A near six-point jump on LiveCodeBench, a benchmark designed to resist training-data contamination, suggests the improvement is genuine generalization rather than memorized solutions.

For teams already using Qwen 3.5 as a coding backend, the upgrade is low risk: the model speaks the same OpenAI-compatible API and keeps the same context budget, so you mostly inherit the higher scores without reworking your prompts.

Reasoning and math: no regression

A persistent claim circulating in forums and a few third-party writeups is that Qwen 3.6 regressed on reasoning, specifically GPQA. The official model card does not support that. GPQA Diamond improves from 84.2 to 86.0 and AIME 2026 improves from 91.0 to 92.7. Both move in the right direction.

The confusion likely comes from comparing different model sizes or different Qwen 3.6 variants, since several public tests benchmarked smaller or differently quantized builds rather than the 35B-A3B release. When you hold the parameter class constant and read the official numbers, Qwen 3.6 35B-A3B does not give up ground on reasoning to win on agentic tasks. It improves on both.

Same architecture, same efficiency

Qwen 3.6 keeps the 35B-A3B mixture-of-experts design: 35 billion total parameters with roughly 3 billion active per token, so per-token compute stays close to Qwen 3.5 on the same hardware. The generational jump comes from training rather than a heavier architecture, which is why scores rise without a larger active parameter count or a bigger memory footprint. For anyone budgeting GPU memory, that continuity is the point: you do not pay more per token to get the better model.

It ships under Apache 2.0 (released April 16, 2026) for unrestricted commercial use, which keeps it viable for closed, revenue-generating products without licensing friction. The available models reference lists the exact deployment specifications.

Context window and thinking mode

Qwen 3.6 35B-A3B supports a 262,144-token native context that extends to roughly 1,010,000 tokens with YaRN scaling. That headroom matters for long agent transcripts, large codebases loaded into a single prompt, and retrieval pipelines that stuff many documents into context. It also includes native thinking mode, where the model spends extra tokens reasoning before it answers, which is most useful on the multi-step coding and math tasks that drive the benchmark gains above. The direct versus thinking guide covers when to enable it, since thinking mode trades latency and token cost for accuracy and is rarely worth it on simple lookups.

Should you upgrade from Qwen 3.5?

For agent and tool-use workloads, the answer is straightforward: yes. The double-digit Terminal-Bench and MCPMark gains translate directly into fewer failed tool calls and shorter agent runs, and the migration is close to free because the API surface is unchanged.

For coding backends, upgrading is also worthwhile. A 3.4-point SWE-bench gain and a 5.8-point LiveCodeBench gain are large enough to notice in real pull-request quality, and there is no regression elsewhere to weigh against them.

For pure chat, summarization, or simple question answering, the gains are smaller. Both releases already score well on reasoning, so if your workload never touches tools, agents, or heavy coding, the upgrade is nice to have rather than urgent. The deciding factor is usually whether your application runs the model inside a loop or just calls it once per request.

Running Qwen 3.6 in production

Qwen 3.6 35B-A3B exposes an OpenAI-compatible API, so most applications migrate by changing the base URL and model name rather than rewriting integration code. Tessera serves it on dedicated GPUs in EU and LATAM regions at a flat monthly price rather than per-token billing, which keeps inference cost predictable for high-volume coding and agentic workloads and keeps data inside the chosen region. That pricing model is a good fit precisely for the agent loops where Qwen 3.6 shines, since those workloads burn many tokens per task and are the most exposed to metered-API bill spikes. The private AI pricing guide breaks down what those flat plans cost. If you are coming from a metered API, see how EU data residency hosting works, and the full Qwen 3.6 35B benchmarks for the complete table and competitor comparison.

Bottom line

Qwen 3.6 35B-A3B is a clean win over the 3.5 release of the same size. It improves on every benchmark Alibaba publishes, posts double-digit gains on the agentic and tool-use tests that decide whether an agent loop is usable, and does it without growing the active parameter count or shrinking the context window. The reasoning regression story does not hold up against the official numbers. If you run agents, coding assistants, or tool-calling pipelines on Qwen 3.5 today, moving to 3.6 is a low-effort, high-return upgrade. If you only use the model for simple one-shot prompts, the gains are real but less pressing.

FAQ

Is Qwen 3.6 better than Qwen 3.5?

Yes. On Alibaba’s official model card, Qwen 3.6 35B-A3B beats the Qwen 3.5 35B-A3B release on every listed benchmark. The biggest gains are in agentic and tool use: Terminal-Bench 2.0 rises from 40.5 to 51.5 and MCPMark from 27.0 to 37.0.

How much better is Qwen 3.6 at coding?

SWE-bench Verified improves from 70.0 to 73.4 and LiveCodeBench v6 from 74.6 to 80.4. Both measure practical coding ability rather than synthetic puzzles, and the LiveCodeBench test is built to resist training-data contamination.

Did Qwen 3.6 regress on any benchmark versus 3.5?

No. On the official model-card suite there is no regression. GPQA Diamond improves from 84.2 to 86.0 and AIME 2026 from 91.0 to 92.7, contrary to some third-party claims of a reasoning regression that compared different model sizes or variants.

Can I run Qwen 3.6 with an OpenAI-compatible API?

Yes. Qwen 3.6 35B-A3B serves through OpenAI-compatible endpoints, so existing application code works by pointing it at a new base URL and model name.

Is Qwen 3.6 open source?

Yes. Alibaba released it under Apache 2.0 on April 16, 2026, allowing unrestricted commercial use.