An 87.2 GiB GGUF of Qwen3.8-Flash-Next hits roughly 8.3 tok/s decode on CPU-only llama.cpp and ~109 tok/s with full VRAM, with GPU-placed embeddings giving a large boost at long context.
The September release adds child devices, a Matter network map, tile-card lighting and climate controls, and OpenStreetMap vector tiles, building on the 2026.8 llama.cpp and Ollama integrations that let local models drive devices with no cloud round-trip.
A Snapdragon X2 Elite AI PC with an 18-core Oryon CPU, Adreno GPU and Hexagon NPU designed to run models locally on-device, available for enterprise purchase from September 20, with a HUMAIN OS variant planned for 2027.
Sink-free visual token pruning that reduces the inference overhead of multimodal LLMs on long visual token sequences, tackling the cost of vision-centric and text-guided pruning strategies.
Details the Qwen3.8-Flash-Next architecture — a 125B sparse MoE (6B activated) with n-gram embedding tables held off-accelerator and hybrid Gated DeltaNet/global attention — beating the 397B model on most benchmarks at a third the activated parameters and roughly a ninth of training FLOPs.
Compresses conversational histories and long documents into continuous memory tokens a frozen language model can read directly, cutting context memory footprint versus human-readable text or decoded images.
The Linear Reusable Neural Bases Architecture (LRNBA) improves parameter efficiency to ease the memory-cost bottleneck in large-model training and inference.
A 305B sparse MoE with vision released under MIT license, the first downloadable checkpoint in the V4 family. Ships in FP8 (~168GB) with tokenizer and minimal PyTorch inference code; the API launched August 21 and weights followed ten days later.
Alibaba released the 35B total / ~3B active multimodal sparse MoE under Apache 2.0, focused on agentic coding with a thinking-preservation feature and day-one support across vLLM, SGLang, Ollama and Unsloth GGUFs (~22-23GB RAM to run locally).
Part of the first open-weight Qwen3.6 release, the 27B dense model scores strong agentic coding and reasoning results and is available in HF Transformers, vLLM, SGLang plus Unsloth GGUF/MTP and MLX-8bit quantizations for local hardware.
A self-hosted AGPL engine that spins up a sandboxed coding agent with repos, tools and credentials from a single prompt. Supports Claude Code, Codex, Gemini CLI and OpenCode runtimes, attaches MCP servers, and keeps conversations in your own Postgres.
A free, MIT-licensed Mac app where a local coding agent lives: drives a real browser signed in as you, streams a live iPhone simulator, keeps chats across restarts, and runs fully on-device with no account and no telemetry.
MCP-shaped tunneling that gives a local stdio MCP server a public URL in one command, with OAuth 2.1-gated endpoints by default and a self-hostable relay. Bridges stdio to Streamable HTTP so remote clients like ChatGPT, Claude and Cursor can reach a server on your laptop.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
