Releases·Model Releasefeatured
Sber's ai-sage releases GigaChat-3.5-Reasoning: 432B-A28B MoE under MIT license
AI-sage released GigaChat-3.5 Reasoning, a 432B-A28B MoE with Gated DeltaNet for long-context efficiency. Domain experts (code, math, general) were trained with CISPO and merged into one model via on-policy distillation. The team reports performance near DeepSeek V4 Flash Preview while emitting 37% fewer reasoning tokens. Weights are public on Hugging Face under MIT, with a hosted demo at giga.chat.
217 engagement1 sourceThu, Sep 10, 2026, 12:19 PM
AI Overview
- GigaChat-3.5-Reasoning is a 432B-A28B MoE with Gated DeltaNet for long-context efficiency, released on Hugging Face under MIT
- Domain experts (code, math, general) were trained with CISPO, then merged via on-policy distillation into one model
- In ai-sage evals the model lands near DeepSeek V4 Flash Preview while using 37% fewer reasoning tokens
Related
Releases·Model Release
AI9Stars releases G9v3-3B open-weight language model
AI9Stars has released G9v3-3B, a 3-billion-parameter open-weight language model under the Apache 2.0 license. The model is designed for strong reasoning in a lightweight package, suitable as a general AI assistant rather than a specialized agent. It is available on GitHub and Hugging Face.
AI9Stars has released G9v3-3B, a 3-billion-parameter open-weight language model under the Apache 2.0 license. The model is designed for strong reasoning in a lightweight package, suitable as a general AI assistant rather than a specialized agent. It is available on GitHub and Hugging Face.
54 engagement1 sourceredditThu, Jul 23, 2026, 07:01 AM
Releases·Model Release
AI9Stars releases G9v3-39A5B, an open-weights MoE reasoning model
AI9Stars released G9v3-39A5B, an open-weights language model with 39B total parameters and 5 active experts, designed for stronger reasoning than its predecessor G9v3-3B. Released under Apache 2.0, it targets assistant, coding, tool-use, and reasoning tasks with Think/No Think modes.
AI9Stars released G9v3-39A5B, an open-weights language model with 39B total parameters and 5 active experts, designed for stronger reasoning than its predecessor G9v3-3B. Released under Apache 2.0, it targets assistant, coding, tool-use, and reasoning tasks with Think/No Think modes.
100 engagement2 sourcesredditMon, Aug 3, 2026, 10:15 AM
Releases·Model Releasefeatured
IBM releases Granite 4.2 family with 3B, 8B, 30B reasoning models
IBM launched Granite 4.2, a family of open-weight, decoder-only reasoning LLMs in 3B, 8B, and 30B sizes, pre-trained from scratch on ~15T tokens. The 8B and 30B variants undergo agentic reinforcement learning for tool use, and all support a 128K native context window.
IBM launched Granite 4.2, a family of open-weight, decoder-only reasoning LLMs in 3B, 8B, and 30B sizes, pre-trained from scratch on ~15T tokens. The 8B and 30B variants undergo agentic reinforcement learning for tool use, and all support a 128K native context window.
120480 engagement3 sourcesreddit, rssTue, Aug 25, 2026, 03:10 PM
Developments·Paper
HLA adds query-dependent chunk mixing to Gated DeltaNet, lifting LongBench-V2 by up to 5.57 points on Qwen3.5
An arXiv paper introduces Hybrid Linear Attention (HLA), a query-dependent chunk-level attention mechanism built on Gated DeltaNet that routes historical states adaptively per query. Evaluated on Qwen3.5 models from 0.8B to 9B, HLA improves long-context benchmarks like LongBench-V2 and RULER over native GDN and fixed chunk mixing.
An arXiv paper introduces Hybrid Linear Attention (HLA), a query-dependent chunk-level attention mechanism built on Gated DeltaNet that routes historical states adaptively per query. Evaluated on Qwen3.5 models from 0.8B to 9B, HLA improves long-context benchmarks like LongBench-V2 and RULER over native GDN and fixed chunk mixing.
0 engagement1 sourcearxivMon, Oct 5, 2026, 05:49 AM
Releases·Tool Release
Unsloth ships large performance release: Vulkan on AMD, MLX speedups, PyTorch 2.11
Unsloth added local-run support for Qwen3.8-Flash-Next and GLM-5.3-Flash with GGUF weights on Hugging Face, then shipped follow-ups enabling multi-token prediction for up to 2x faster inference and a major performance/bugfix release. The latest release switches AMD to Vulkan by default for a 20% prefill/decoding boost over ROCm and includes 200+ bug fixes. Practitioners gain faster local GGUF inference, one-click Hermes support, and MLX fine-tuning on Apple Silicon.
Unsloth added local-run support for Qwen3.8-Flash-Next and GLM-5.3-Flash with GGUF weights on Hugging Face, then shipped follow-ups enabling multi-token prediction for up to 2x faster inference and a major performance/bugfix release. The latest release switches AMD to Vulkan by default for a 20% prefill/decoding boost over ROCm and includes 200+ bug fixes. Practitioners gain faster local GGUF inference, one-click Hermes support, and MLX fine-tuning on Apple Silicon.
Updated Wed, Sep 9, 2026, 03:18 PM — September 8 release defaults AMD to Vulkan for 20% perf boost, fixes AMD gibberish issues, and adds 200+ bug fixes
1601.8k engagement7 sourcesgithub, xThu, Aug 27, 2026, 01:09 PM