Qwen3.6-35B-A3B
A Mixture-of-Experts (MoE) model in Alibaba's Qwen series, with approximately 36 billion total parameters and approximately 3.5 billion activated parameters per forward pass. Its sparse activation architecture results in substantially lower inference costs compared with dense models of equivalent total parameter count. The model supports 4-bit quantization for local deployment.
Core Facts
Timeline (last 90 days)
MIRA's current results are based on the specific foundation model Qwen3.6-35B-A3B, and the generality of MIRA-Harness and MIRA-Data still needs to be verified on more models.
Drex has open weights but uses the Open RAIL-M license with usage restrictions, so it is not open source as defined by OSI; it offers both hosted and self-hosted channels; it is based on Qwen3.6-35B-A3B MoE, with 35B total parameters and about 3B activated; self-hosted default context is 16K with a maximum of 128K, and hosted is 128K.
When run on Qwen3.6-35B-A3B with a 50% expert pruning ratio, SlimWise increases decoding throughput by up to 1.81x with minimal accuracy loss.
For building local agent tasks (such as automated script orchestration, browser operation, file management), the speed advantage of LFM2.5-24B-A2B is more suitable; for code assistance tasks, Qwen3.6-35B-A3B is more suitable
LFM2.5-24B-A2B and Qwen3.6-35B-A3B are both open-source, locally deployable MoE architecture models, with hardware requirements within the range of high-end consumer devices
Data released by Liquid AI shows that LFM2.5-24B-A2B beats Qwen3.6-35B-A3B on tool-use benchmarks
LFM2.5-24B-A2B surpassed Qwen3.6-35B-A3B on five of eight tool-use benchmarks.
On the AutomationBench evaluation, LFM2.5-24B-A2B scored 27.69%, higher than Qwen3.6-35B-A3B's 8.72%, Gemma-4-26B-A4B's 5.46% and Nemotron-3.5-Lightning's 2.61%.
On the SWE-bench Verified evaluation, LFM2.5-24B-A2B scored 51.00%, trailing Qwen3.6-35B-A3B's 66.20% and Claude Sonnet 5.5's 55.2%.
LFM2.5-24B-A2B outperforms Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning (30B-A3B) on multiple instruction-following and agentic benchmarks, but lags on knowledge-intensive benchmarks such as SWE-Bench Verified and GPQA Diamond.
7 more timeline events
All Facts (17)
Qwen3.6 35B-A3B是混合专家架构（MoE），总参数量35B，每次推理实际激活参数量约3B
70%UnverifiedMIRA's current results are based on the specific foundation model Qwen3.6-35B-A3B, and the generality of MIRA-Harness and MIRA-Data still needs to be verified on more models.
50%UnverifiedDrex has open weights but uses the Open RAIL-M license with usage restrictions, so it is not open source as defined by OSI; it offers both hosted and self-hosted channels; it is based on Qwen3.6-35B-A3B MoE, with 35B total parameters and about 3B activated; self-hosted default context is 16K with a maximum of 128K, and hosted is 128K.
50%UnverifiedWhen run on Qwen3.6-35B-A3B with a 50% expert pruning ratio, SlimWise increases decoding throughput by up to 1.81x with minimal accuracy loss.
50%UnverifiedFor building local agent tasks (such as automated script orchestration, browser operation, file management), the speed advantage of LFM2.5-24B-A2B is more suitable; for code assistance tasks, Qwen3.6-35B-A3B is more suitable
50%UnverifiedData released by Liquid AI shows that LFM2.5-24B-A2B beats Qwen3.6-35B-A3B on tool-use benchmarks
50%UnverifiedLFM2.5-24B-A2B and Qwen3.6-35B-A3B are both open-source, locally deployable MoE architecture models, with hardware requirements within the range of high-end consumer devices
50%UnverifiedLFM2.5-24B-A2B surpassed Qwen3.6-35B-A3B on five of eight tool-use benchmarks.
50%UnverifiedOn the AutomationBench evaluation, LFM2.5-24B-A2B scored 27.69%, higher than Qwen3.6-35B-A3B's 8.72%, Gemma-4-26B-A4B's 5.46% and Nemotron-3.5-Lightning's 2.61%.
50%UnverifiedOn the SWE-bench Verified evaluation, LFM2.5-24B-A2B scored 51.00%, trailing Qwen3.6-35B-A3B's 66.20% and Claude Sonnet 5.5's 55.2%.
50%UnverifiedLFM2.5-24B-A2B outperforms Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning (30B-A3B) on multiple instruction-following and agentic benchmarks, but lags on knowledge-intensive benchmarks such as SWE-Bench Verified and GPQA Diamond.
50%UnverifiedLFM2.5-24B-A2B scored 48.09 on the LiveClaw benchmark.
50%UnverifiedIn the task of determining the type of user utterance request, Qwen3.6-35B-A3B achieved an accuracy of 92.6%, with a judgment time of 306ms.
50%UnverifiedOn "continuation of previous operation" type utterances, Qwen3.6-35B-A3B achieved an accuracy of 86.3%, while the 4B and 8B models achieved only 70.4%.
50%UnverifiedThe Agent's default self-hosted configuration uses the Qwen3.6-35B-A3B model with 4-bit quantization deployed via sglang.
50%UnverifiedThe developer uses the Qwen3.6-35B-A3B model: it has 35B weights in total, uses only 3B weights when processing each word, has a computational cost comparable to ordinary 4B and 8B models, and runs on their own gateway (vLLM).
50%UnverifiedThe locally deployed Qwen3.6-35B-A3B model once caught real bugs in code originally written by Claude Sonnet/Opus, and the author emphasized this was not a benchmark test but happened by chance.
50%