Qwen 3.6 35B A3B

Qwen 3.6 35B A3B is a mid-sized language model in the Qwen family that carries forward the mixture-of-experts approach used in the previous generation, keeping a roughly three-billion active parameter footprint within a larger 35-billion parameter total, which is intended to balance capability with inference efficiency. The release is positioned as a generational step within the same size tier rather than a scaling shift, and the design intent appears to favor strong agentic reasoning, code work, and tool orchestration at a moderate compute cost.

Third-party reporting on the model's official benchmark card highlights notable gains in coding and agent-style tasks, with a SWE-bench Verified score of 73.4 versus 70.0 for the prior generation, a jump from 40.5 to 51.5 on Terminal-Bench 2.0, and an increase from 27.0 to 37.0 on MCPMark, alongside an 80.4 reading on LiveCodeBench v6. These same sources note no regressions across the reported suites, with reasoning and math also improving, suggesting practical value for developer workflows that lean on multi-step tool use and code generation.

Venice AIqwen3-6-35b-a3bqwen

Open provider page Provider docs

Quick Info

Provider: Venice AI
Model key: qwen3-6-35b-a3b
Release date: Jul 20, 2026
Last updated: Jul 22, 2026
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.10
Output token cost: $1.00

Limits

Output tokens: 65,536 tokens
Context window: 256,000 tokens

Transparent token rates

Compare Qwen 3.6 35B A3B pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about Qwen 3.6 35B A3B

Venice AI

CoverageBenchmark

Qwen 3.6 35B-A3B coding benchmark scores include SWE-bench Verified 73.4, Terminal-Bench 2.0 51.5, LiveCodeBench v6 80.4, MCPMark 37.0, GPQA Diamond 86.0, and AIME 2026 92.7. These figures are sourced from Alibaba's official model card on Hugging Face, and the article also flags community-reported SWE-bench Multilingua Generation-over-generation comparison shows notable gains over Qwen 3.5 35B-A3B: Terminal-Bench 2.0 rises by 11 points and MCPMark by 10 points, both reflecting improvements in multi-step tool use. SWE-bench Verified climbs 3.4 points to 73.4, and the article frames the model as one of the strongest mid-sized open-weig

Venice AI

CoverageBenchmark

A third-party GGUF quality benchmark published April 19, 2026 evaluated 64 quantized builds of Qwen 3.6 35B A3B from six uploaders — unsloth, bartowski, lmstudio-community, ggml-org, mudler, and AesSedai — against a BF16 reference using KL divergence across 250,000 tokens of coding, chat, tool calling, science, non-Lat The benchmark identified unsloth as dominating the Pareto frontier with 14 of 26 positions, while AesSedai filled remaining size gaps, and both lmstudio-community and mudler (APEX) never appeared on the frontier. A noteworthy anomaly surfaced: tool calling was the worst-performing category at Q8_0 with a KL of 0.177, w