more pics from the Tenstorrent lab of my > 4x QuietBox Blackhole Cluster by the time i receive them in a few weeks > 7 NEW MoE architectures will be SUPPORTED specs of each QuietBox: > EPYC 8124P (16C/32T, 3 GHz)
> 512 GB DDR5-4800 ECC RDIMM > 4x Blackhole p150c: 128GB VRAM
LLMS
-
Tenstorrent QuietBox Blackhole Cluster specs and MoE support
By
–
-
Qwen Audio Model Now Available: 70GB Weights for Local Deployment
By
–
This is really cool – you can try it out by signing into https://
chat.qwen.ai (with Google or GitHub) and selecting the audio icon The model weights are only ~70GB and that's before quantizing them down, so this one is going to be reasonably accessible to run locally -
Teaching LLMs Logical Chain-of-Thought for Symbolic Planning
By
–
Teaching LLMs to Plan: Logical Chain-of-Thought Instruction Tuning for Symbolic Planning Verma et al.: https://
arxiv.org/abs/2509.13351 #ArtificialIntelligence #DeepLearning #MachineLearning -
ChatLLM Desktop UX Improvements and Feature Development
By
–
UX Improvements – We are working on improving our UX even more Please comment specifically on how we can improve the ChatLLM UX on desktop. We are aware that some features are missing on the app and we are working to improve them
-
Qwen Releases S Tier Lab Model with Advanced Features
By
–
Qwen are playing to win S Tier lab awesome model in both size and features
-
DeepSeek-V3.1-Terminus Released on Hugging Face with Improvements
By
–
DeepSeek-V3.1-Terminus is out on Hugging Face
— AK (@_akhaliq) 22 septembre 2025
This update maintains the model's original capabilities while addressing issues reported by users, including:
Language consistency: Reducing instances of mixed Chinese-English text and occasional abnormal characters
Agent… pic.twitter.com/bw9AtNHFUrDeepSeek-V3.1-Terminus is out on Hugging Face This update maintains the model's original capabilities while addressing issues reported by users, including: Language consistency: Reducing instances of mixed Chinese-English text and occasional abnormal characters Agent
-
Controlling AI dialogue consistency and speaker attribution challenges
By
–
even on runway, if the dialogue went back and forth too much, it didn't keep consistency on who speaks, and was hard to control
— Yohei (@yoheinakajima) 22 septembre 2025
(example video here messes this up) pic.twitter.com/IMt6vgpyhReven on runway, if the dialogue went back and forth too much, it didn't keep consistency on who speaks, and was hard to control (example video here messes this up)
-
Open-source AI models rapidly closing the gap with proprietary solutions
By
–
Open-source models are catching up fast: •Wan 2.2 & SeeDance → strong video models
•Flux & Dreamina → top-tier images
•Qwen & Deepseek → just shy of SOTA LLMs
•Kimi → impressive agentic loops You can actually get serious work done with open source, and most are -
DeepSeek-V3.1-Terminus Update: Chat and Reasoner Models Upgraded
By
–
Both deepseek-chat and deepseek-reasoner have been upgraded to DeepSeek-V3.1-Terminus. deepseek-chat corresponds to DeepSeek-V3.1-Terminus's non-thinking mode, while deepseek-reasoner corresponds to its thinking mode.
-
Deepseek V3.1 Terminus: Improved Language Consistency
By
–
Deepseek released Deepseek V3.1 Terminus with improved language consistency and minor agent updates for coding and search. Less CN replies for English queries