Qwen3.6-35B-A3B

2026-10-21

RESEARCH · 1 SOURCE · arXiv cs.AI

Navalia framework uses verifiable task synthesis and long-horizon post-training to improve Cowork Agents

Navalia is a framework that builds verifiable analytical tasks from corporate financial reports and applies long-horizon post-training (SFT and RL) to improve Cowork Agent performance. Initialized from Qwen3.6-35B-A3B, the approach uses semantic evidence selection, executable computation chains, and independent verification to generate training trajectories and yields consistent gains across five benchmarks (OfficeQA Pro, FinSearchComp, Finance Agent Benchmark, APEX-Agents-AA, Workspace-Bench-Lite).

7.0

2026-10-20

CODING · 1 SOURCE · Hugging Face

PyTorch 2.15 brings FP8, faster LLM decoding and new Metal MPS kernels for Macs

PyTorch 2.15 (Hugging Face contributors) improves the MPS backend on Apple Silicon: adds torch.float8_e4m3fn (FP8) support with F.scaled_mm, faster single‑row linear and grouped matmul paths that speed Mixture‑of‑Experts decoding 5–6.5× (e.g., Qwen3.6‑35B‑A3B from 9.4→61.3 tok/s on an M3 Ultra), an ~8× faster softmax for sampling, and many new Metal kernels that cut driver memory for some video decoders (Wan2.2) from ~270 GiB to ~8 GiB. The release also raises platform requirements to macOS 15+ and Python 3.11+, and previews further MPSGraph→Metal migrations in 2.16.

7.0

2026-10-13

MODELS · 1 SOURCE · Hugging Face

LFM2.5-24B-A2B: 24B-parameter MoE agent model (2B active) for fast private on-device reasoning — open weights

LFM2.5-24B-A2B is a 24B-parameter MoE reasoning model that activates only 2B parameters per token, is reasoning‑only (emits explicit chains of thought), and is optimized for continuous, private agent use on-device. It expands context to 131K tokens, doubles vocabulary to 128K, reports high edge and GPU throughput (e.g., 184 tok/s on Apple M5 Max and >21K tok/s on an NVIDIA B300), is competitive with larger models on agentic and tool‑use benchmarks, and is available as open weights on Hugging Face with day‑one support in major inference runtimes (llama.cpp, MLX, vLLM, SGLang, ONNX).

8.0

2026-09-24

RESEARCH · 1 SOURCE · arXiv cs.AI

VHD-Play paper: generate solved, verifiable agentic environments and train LLM agents

arXiv:2609.27321v1 presents VHD-Play, a pipeline that samples and solves mathematical mechanisms first and then renders them as corpus-grounded, stateful environments with inherited executable dynamics and trajectory scoring; the authors say it produced 3,300 diverse agentic environments at a cost of a few cents each. Training Qwen3.6-35B-A3B on three families reportedly raised its mean agentic score from 0.204 to 0.815 on a five-family diagnostic, generalized to held-out and eight unseen mechanism families, and improved external benchmarks (on E-Commerce Bench the checkpoint completed every run without bankruptcy and exceeded Qwen3.7-Max).

8.0