AI MODEL MONITOR
AI Model Release Monitor — Daily Digest (2026-10-04)
2026-10-04
9 new model release(s) found.
CHINA / LOCAL MODELS
Index-Translate family (Bilibili IndexTeam)
Hugging Face (IndexTeam)
Bilibili IndexTeam’s open-weight translation family covers text translation across 150 languages and follows instructions for terminology, style, formatting, and length. The newly posted 35B-A3B preview complements 2B/9B text models; related family models extend into speech translation, controlled dubbing, and long documents.
Hardware: 2B: ~5GB VRAM at full precision (FP16/BF16), ~2GB at INT8, ~1GB at 4-bit (GGUF/AWQ) quantization. Runs comfortably on a single consumer GPU (8-12GB) or even CPU/laptop RAM at Q4. 9B: ~21GB VRAM at full precision (FP16/BF16), ~11GB at INT8, ~5GB at 4-bit (GGUF/AWQ) quantization. Runs comfortably on a single consumer GPU (8-12GB) or even CPU/laptop RAM at Q4. 35B-A3B preview: ~79GB VRAM at full precision (FP16/BF16), ~40GB at INT8, ~20GB 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.)
AdvancedMathBench-AutoVerifier
(InternLM (Shanghai AI Laboratory))
Hugging Face (InternLM)
A learned verifier for natural-language mathematical proofs that grades a proof, points to the earliest incorrect step, and explains errors; it ships with InternLM’s AdvancedMathBench/ProverBench work. It is a research grader rather than a formal proof checker.
Hardware: ~79GB VRAM at full precision (FP16/BF16), ~40GB at INT8, ~20GB at 4-bit (GGUF/AWQ) quantization. Needs a workstation-class GPU or multi-GPU consumer setup (2-3x 24GB cards) at Q4.
OTHER RELEASES
Strands Decider 2B
(AWS Strands Labs)
Strands Agents
An open decision model for agent routing, tool selection, triage, and scoring: it ranks developer-supplied choices and returns confidence rather than generating prose. AWS released the weights, training data, and scripts; the team reports local inference on CPU/GPU, including about 115ms median latency on an RTX 3090.
Hardware: HF API did not return safetensors.total; estimate uses the official ~1.9B base size: ~4GB VRAM at full precision (FP16/BF16), ~2GB at INT8, ~1GB at 4-bit (GGUF/AWQ) quantization. Runs comfortably on a single consumer GPU (8-12GB) or even CPU/laptop RAM at Q4.
Kolibri-1
(Aleph Alpha)
Hugging Face
Aleph Alpha’s open-weight English/German reasoning MoE supports long-context document work, retrieval, coding, and tool calling, with a one-million-token validated context extension (262K native training length). The full 78.1B weights—not just the 3.46B active per token—must be loaded.
Hardware: ~172GB VRAM at full precision (FP16/BF16), ~86GB at INT8, ~43GB 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.) Official FP8 checkpoint is about 78GB; the model card lists a minimum of 2× A100 80GB or 1× H200/B200/B300. BF16 weights require about 156GB before serving overhead.
AstaBrief-8B
(Allen Institute for AI (Ai2))
Hugging Face Blog
Ai2’s open-weight scientific-report model turns a research question plus retrieved paper excerpts into a cited report in one pass. Ai2 also released training data and a local-PDF workflow for institutions that need to keep research materials on their own infrastructure.
Hardware: HF API did not return safetensors.total; estimate uses the model card’s 8B size: ~18GB VRAM at full precision (FP16/BF16), ~9GB at INT8, ~4GB at 4-bit (GGUF/AWQ) quantization. Runs comfortably on a single consumer GPU (8-12GB) or even CPU/laptop RAM at Q4.
Cloudflare Clef / Clef-Flash
(Cloudflare)
Cloudflare Blog
These multimodal decision models score typed options from text, JSON, images, or video instead of generating free-form text. Clef (27.36B) prioritizes precision, while Clef-Flash (9.41B) is the lower-latency variant; both are open-weight and Jev/SystemOne API-compatible.
Hardware: Clef: ~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. Clef-Flash: ~21GB VRAM at full precision (FP16/BF16), ~10GB at INT8, ~5GB at 4-bit (GGUF/AWQ) quantization. Runs comfortably on a single consumer GPU (8-12GB) or even CPU/laptop RAM at Q4.
Gemini 4 Argon
(Google DeepMind)
Google Blog
Google announced a frontier hosted model for long-horizon software engineering, enterprise knowledge work, and cybersecurity, with a one-million-token output limit. Initial access is limited to trusted cyber defenders through Fairwind and Google internal teams; broader availability is planned.
Hardware: Limited-access hosted model; weights and parameter count are not public, so no local hardware requirement is available.
MAI-Transcribe-2-Streaming
(Microsoft AI)
Microsoft AI
Microsoft’s real-time speech-to-text model streams partial and final transcripts in 60 languages with continuous language detection; the first partial arrives in roughly 120ms. It is available through Microsoft Foundry and the MAI Playground, with no public weights or parameter count.
Hardware: Hosted API only; model weights and parameter count are not public, so no local hardware requirement is available.
MAI-Voice-2.1 / MAI-Voice-2.1-Flash
(Microsoft AI)
Microsoft AI
Microsoft’s two hosted text-to-speech releases support 23 languages and 26 locales while maintaining a consistent voice across languages. Flash targets latency-sensitive, high-volume use; neither model has public weights or a disclosed parameter count.
Hardware: Hosted APIs only; model weights and parameter counts are not public, so no local hardware requirement is available.
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 — Show us the task nobody wants to do. We'll show you what AI can do with it.