AI MODEL MONITOR
AI Model Release Monitor — Daily Digest (2026-09-29)
2026-09-29
6 new model release(s) found.
CHINA / LOCAL MODELS
Qwen-Audio-3.1-Realtime (Alibaba Qwen)
MarkTechPost
Full-duplex real-time voice model for agents that can reason, call tools, and decide when to speak, with speech/text input and output. Reported as available through QwenCloud; no downloadable weights or parameter count were published.
Hardware: Hosted API only; no local hardware requirement published.
IQuest-Q1
(IQuest)
Hugging Face
Open-weight MoE for agentic coding, software engineering, reasoning, and multi-step tool use, with a 512K-token context. The model card describes it as early-stage and text-only.
Hardware: ~705GB VRAM at full precision (FP16/BF16), ~352GB at INT8, ~176GB at 4-bit (GGUF/AWQ) quantization. Effectively datacenter-only – multiple 80GB+ GPUs even at 4-bit quantization; not practical on consumer hardware. (MoE: only ~15B active per token, so inference speed is much faster than the total size implies, but full weights still need to be loaded.)
OTHER RELEASES
Claude Sonnet 5.5 (Anthropic)
Anthropic
Hosted model aimed at everyday knowledge work, coding agents, and bug fixing; Anthropic says it is over 30% faster than Sonnet 5 and can cost up to 30% less per task. Available through Claude and major cloud platforms.
Hardware: Hosted service; no local hardware requirement or model weights published.
Holo4-27B (H Company)
Hugging Face Blog
Dense agentic model for computer use across GUIs, code, MCP, and APIs; trained for long, multi-step business workflows. Open weights are available, but commercial use is not permitted under its listed license.
Hardware: ~60GB VRAM at full precision (FP16/BF16), ~30GB at INT8, ~15GB at 4-bit (GGUF/AWQ) quantization. Fits on a single high-end consumer GPU (RTX 4090/5090, 16-24GB) at Q4.
Holo4-35B-A3B (H Company)
Hugging Face Blog
Open-weight MoE agent for computer use, coding, MCP, and API tools across desktops, web, and Android. Apache-2.0 licensed for self-hosting; only about 3B parameters activate per token, but all weights must be resident.
Hardware: ~77GB VRAM at full precision (FP16/BF16), ~39GB at INT8, ~19GB at 4-bit (GGUF/AWQ) quantization. Needs a workstation-class GPU or multi-GPU consumer setup (2-3x 24GB cards) at Q4. (MoE: only ~3B active per token, so inference speed is much faster than the total size implies, but full weights still need to be loaded.)
Holotron4 Nano (H Company)
Hugging Face Blog
Updated Nemotron-based generalist agent for GUI workflows, MCP/API environments, and coding sandboxes. H Company says its post-training stack improves the underlying Nemotron 3 Nano Omni model.
Hardware: ~73GB VRAM at full precision (FP16/BF16), ~36GB at INT8, ~18GB at 4-bit (GGUF/AWQ) quantization. Needs a workstation-class GPU or multi-GPU consumer setup (2-3x 24GB cards) at Q4. (MoE: only ~3B active per token, so inference speed is much faster than the total size implies, but full weights still need to be loaded.)
Full browsable, searchable model page
Source: https://drive.google.com/file/d/1NpL8_FQv2YM5fM2x8N9A5w8jgfi3xrjQ/view?usp=sharing
Source: https://drive.google.com/file/d/1NpL8_FQv2YM5fM2x8N9A5w8jgfi3xrjQ/view?usp=sharing
RAW1 AI — Stop doing manually what technology can do automatically.