A hands-on build and run of the Swift-based LLM inference server SwiftLM on an M1 Max 64GB. Covers Qwen3.6-35B-A3B and Qwen3.5-122B-A10B, with the same BST, BBS, and persona tests used in the existing Ollama and MLX-lm write-ups.
DeepSeek V4 Preview ships V4-Pro (1.6T/49B active) and V4-Flash (284B/13B active) as open weights under MIT, both with 1M context. CSA+HCA hybrid attention, mHC, and the Muon optimizer cut per-token FLOPs at 1M tokens to 27% of V3.2. Day-one API and chat.deepseek.com mode switch covered.
Two open-weight Chinese MoEs landed within 24 hours: Ant Ling-2.6-flash (104B/7.4B active, 7x token-efficiency claim) and Tencent Hy3-preview (295B/21B active, frontier-tier open weights). Specs, licenses, and how they line up against DeepSeek-V3 and GLM-4.5.
Xiaomi launched two MiMo-V2.5 models at once. MiMo-V2.5-Pro hits SWE-bench Pro 57.2, Claw-Eval 63.8, and τ3-Bench 72.9 — frontier-tier — while MiMo-V2.5 brings native omnimodality plus a 1M context. Both are API-only for now; open weights are promised but unscheduled.
Tried Qwen3.6-27B on both Ollama and MLX. Ollama couldn't load the VL-projector-embedded GGUF, MLX ran it at 11 tok/s. On the side, running 35B-A3B under MLX was roughly 2× faster than the Ollama GGUF. Also had both models build a BBS to gauge intent handling.
Hands-on Qwen3.6-35B-A3B (23GB 4bit GGUF) on M1 Max 64GB via Ollama 0.20.6. Generation speed stays at 27 tok/s — same as Qwen3.5-35B-A3B — but the same prompt produces 13× more thinking tokens. Multi-turn behavior, persona handling, and a three-tier NSFW probe included.
Alibaba's Qwen3.6-Max-Preview and Moonshot AI's Kimi K2.6 were released within a 24-hour window on April 20–21, 2026. A side-by-side look at specs, benchmarks, distribution, and agent-side features for the two flagships.
Alibaba's Qwen team released Qwen3.6-35B-A3B as open weights. A 40-layer hybrid of Gated DeltaNet, Gated Attention, and MoE hits 73.4 on SWE-bench Verified, 37.0 on MCPMark, and 1397 on QwenWebBench.
Zhipu AI's GLM-5.1 is a 744B MoE (40B active, 200K context, MIT) targeting long-horizon agent tasks. Hits 58.4% SOTA on SWE-Bench Pro (edging out GPT-5.4 and Claude Opus 4.6) and sustains performance across 8-hour sessions with 6,000+ tool calls without degradation.
Benchmarking NII's LLM-jp-4-32B-A3B-thinking on EVO-X2 (Ryzen AI Max+ 395) with ROCm. 62.9 t/s vs Qwen3.5-35B-A3B's 44.7 t/s. Covers thinking control issues, KV cache trade-offs, knowledge cutoff, Japanese quality comparisons, code generation tests, and training data composition.
Google DeepMind has released Gemma 4: four models—31B dense, 26B MoE (A4B), E4B, and E2B—with a 256K context, multimodal input, tool calling, and support for 140 languages.
SwiftLM, an Apple Silicon–only MLX inference server, provides a native Metal implementation of TurboQuant V2+V3 hybrid KV‑cache compression and NVMe SSD expert streaming.
