Nvidia Nemotron Bolt 3.5 74B A9B set for Oct. 15
A Reddit post says Nvidia's Nemotron Bolt 3.5 is a 74B-parameter, 9B-active MoE model quantized to 24GB using 2-bit quantization-aware training for speed.
NVIDIAOpen
Open Models
Archive, page 1
Wednesday 14 October 2026
OrcaRouter compares HiRO-ACEv2 with Qwen 3.8
OrcaRouter contrasts a quietly published repo, HiRO-ACEv2, against the full launch of Qwen 3.8, weighing the two model efforts. The source is largely commentary without detailed benchmark numbers.
OrcaRouterOpen
Leaderboard: China open models trail US by 7 months
A chart from whatmodel.sh, an open-source AI leaderboard, shows Chinese open-weight text models trail US closed frontier models by about seven months, based on 22,113 results across 987 benchmarks from 1,011 sources.
whatmodel.shOpenFreeCode
More: Startup Fortune
FLock.io unveils Qwen-based edge decision model for privacy
FLock.io released an edge decision model built on Qwen aimed at enhanced data privacy, claiming it outperforms larger LLMs on selected tasks.
FLock.ioOpen
More: Korea IT Times
Z.ai GLM 5.3 analysis says post-training is China's new AI edge
BigGo Finance argues Z.ai's GLM 5.3 shows post-training techniques, rather than raw compute, have become China's key competitive advantage in AI.
Z.aiOpen
Maya-2 open-weight Indic voice model claims top performance
An X post announces Maya-2, billed as the world's best Indic voice AI model with fully open weights, claiming it outperforms ElevenLabs v3, Cartesia Sonic 3.5 and Bulbul v3 across Indian languages; free to builders soon.
♥ 12 on X
Maya-2OpenOpen weights
Tuesday 13 October 2026
OrcaRouter benchmarks Qwen-Image-2.1-Turbo locally with BaoCut 3.5
OrcaRouter published first local timing tests running Qwen-Image-2.1-Turbo on a desktop via BaoCut 3.5, and compared it to Z-Image Turbo across two 8-step configurations.
OrcaRouterOpenOpen weightsCode
Comparison of Fish Audio Drama 3 Preview and Qwen Audio 3.0 TTS
OrcaRouter compares two text-to-speech systems, Fish Audio Drama 3 Preview and Qwen Audio 3.0 TTS, on voice quality and expressiveness.
OrcaRouterOpenOpen weights
Jailbroken PS5 runs Qwen 27B at ~20 tok/s
An enthusiast ran a 27 billion parameter Qwen model on a jailbroken PlayStation 5's GPU at about 20 tokens per second, demonstrating local LLM inference on console hardware.
Open
More: Startup Fortune
Built In explainer covers DeepSeek and open-weight AI models
An explainer article describing what DeepSeek is and how open-weight AI models work.
DeepSeekOpen
Liquid AI releases LFM2.5-24B-A2B open-weight model for on-device agents
LFM2.5-24B-A2B is an open-weight Mixture-of-Experts model from Liquid AI with 24B total and about 2B active parameters, designed to run fast, private agent tasks locally on a laptop. Weights are available on Hugging Face.
Liquid AIOpenOpen weights
Kugel 3.5 text-to-speech model launches on Product Hunt
Kugel 3.5 is a text-to-speech model its makers claim outperforms other models on the market. It was shared on Product Hunt; no pricing or availability details were given.
Kugel
OrcaRouter outlines Mistral Large 4 specs, price and missing licence
A post examines Mistral Large 4's specifications, pricing and the absence of a stated licence. It is analysis of an existing model rather than a new release.
Mistral
Reddit post claims Gemma4 26B A4B runs on 14 smart TVs
A Reddit post jokes that Gemma4 26B A4B can be run with only 14 smart TVs. It is a community joke rather than a news item.
Open
Developer converts dense models into Mixture-of-Experts, releases Qwen2.5-1.5B-MoE
An independent developer continued work on converting dense LLMs into Mixture-of-Experts models, releasing Qwen2.5-1.5B-MoE-A0.9B along with GGUF versions for local use.
OpenOpen weightsCode
Developers turn EmbeddingGemma 2 into a speech-to-text model
A project repurposes Google's EmbeddingGemma 2 embedding model into a speech-to-text system, with the write-up and demo shared on Hacker News.
OpenOpen weightsCode
OrcaRouter compares Qwen 3.7 and Qwen 3.8
OrcaRouter published a comparison of Qwen 3.7 and Qwen 3.8 aimed at helping developers decide which model generation to call for their workloads.
QwenOpen
MIT Technology Review weighs open model benefits against risks
MIT Technology Review Explains piece examines the benefits of open-weight models against risks, citing an Anthropic report arguing GLM-5.3 poses a major cyberattack risk.
Open
n+2 lookahead speeds Qwen Flash Next SSD streaming on Mac
A hobbyist implemented an n+2 lookahead prediction for SSD-streaming Qwen Flash Next on a 64GB Mac Mini M5, fetching experts a layer ahead to speed token output.
OpenOpen weights
More: OrcaRouter, Pasquale Pillitteri
Z.ai releases GLM-Legal model for legal tasks
Z.ai's GLM-Legal, built on GLM-5.3-Flash, is a legal AI model scoring 19.2% all-pass on Harvey's Legal Agent Benchmark, nearly 3x its base, at $0.15 per test, by application.
Z.aiClosedAPIWaitlist
DeepSeek and Xiaomi lead China's AI push with 16 new models
Brave New Coin reports DeepSeek and Xiaomi are leading a Chinese AI charge with 16 new models, challenging US tech giants.
DeepSeekOpen
Monday 12 October 2026
Alibaba's Qwen open-sources QwenGyre RL framework
QwenGyre is an elastic online reinforcement learning framework from Alibaba's Qwen team for training extreme-long-horizon coding agents through unmodified black-box harnesses, targeting rollouts that run for hours.
QwenOpenCodeOpen weights
More: Startup Fortune
OpenBMB releases MiniCPM-V 4.7 vision models, first MoE in series
MiniCPM-V 4.7 is the latest in OpenBMB's vision-language series, including a 35B-A3B variant that is its first MoE model, plus a 1B version and GGUF quantizations. It handles single-image, multi-image and video understanding.
OpenBMBOpenOpen weights
More: Hugging Face
Precision Numerics releases Boole-1.7B, a 1-bit Qwen3 model
Boole-1.7B is a 1-bit version of Qwen3-1.7B with 247 MB packed weights, 13.9x smaller than BF16 and scoring 85.9% of the BF16 model across ten tasks. It can be run locally.
Precision NumericsOpenOpen weights
Import AI 476 flags China's open-model edge
The newsletter Import AI's 476th issue, with Clark flagging China's open-model advantage, covers the latest AI developments for 2026.
Ali Qwen launches QwenGyre online RL framework for agent tasks
Alibaba's Qwen launched QwenGyre, an online reinforcement learning framework for long-term agent tasks, per KuCoin.
QwenOpenCode
More: Reddit
DeepSeek V4.1 Flash API draws benchmarks and guides
DeepSeek V4.1 Flash sees setup guides, spec and price breakdowns, plus MiMo-V2.6-Pro comparisons and AMD MI355X AA-AgentPerf results.
DeepSeekOpenAPIOpen weights
More: OrcaRouter, Crypto Briefing
Xiryn launches as open-source personal agent for your computer
Xiryn is an open-source personal AI agent that runs locally on your computer to handle tasks on your behalf.
XirynOpenFreeCode
Reflection AI plans to release America's most powerful open-weight model
Reflection AI says it wants to deliver America's most powerful open-weight AI model. Coverage links the effort to Nvidia's open-weight push. No release date is given.
Reflection AIOpen
More: Yahoo Finance
TII releases Falcon-ASR, a 1.6B Arabic speech recognition model
TII releases Falcon-ASR, a 1.6B-parameter Arabic speech recognition model with an Emirati focus, reporting 20.92% average WER on six test sets. Weights and API are not yet available.
TII
Mistral unveils ML4 open-weight model, claims edge over Chinese rivals
Mistral unveils ML4, an open-weight model that it says gives it an edge over Chinese rivals, as it competes with US labs.
MistralOpenOpen weights
More: WSJ
Sunday 11 October 2026
vLLM Semantic Router releases Decision 3.0, five Apache-2.0 decision models
Decision 3.0 is a family of five open multimodal models from 0.8B to 27B parameters that answer choice, yes/no and score questions about text and images in one call without generating text. Released under Apache-2.0 on Hugging Face.
vLLM Semantic RouterOpenOpen weightsFree
More: Hugging Face
Saturday 10 October 2026
Qwen releases Qwen-Image-2.1 with native RGBA transparency for generation and editing
An image model for text-to-image generation and image editing, with native RGBA transparency support. Interactive demo Space is available on Hugging Face, with weights and blog linked.
Space · gradio · ♥ 407
QwenOpenOpen weightsFree
More: YouTube
Cloudflare releases Clef decision models, including 30B-A3B Clef-omni multimodal model
Open Apache 2.0 models that take a state (text, JSON, images or video) and typed questions, returning probabilities for every option in one forward pass. Clef-omni is on Hugging Face.
Model · image-text-to-text · apache-2.0 · ♥ 2.0k · 13.6k downloads
CloudflareOpenOpen weightsFree
OpenAI enterprise customers can apply existing commitments to Moonshot's Kimi K3 via Baseten
OpenAI enterprise users can apply existing commitments to Moonshot AI's Kimi K3 through Baseten, according to a Reddit post. Details of the arrangement are limited.
Moonshot AIOpenAPI
Vambo AI releases MORENA, a 1.5B open-weight model for 12 African languages
MORENA is a 1.5 billion parameter open-weight language model trained from scratch for 12 African languages. Developers can download the weights and run it themselves.
Vambo AIOpenOpen weights
Nace AI open-sources Drex 1.5, a 9B decision model scoring options
Nace AI has open-sourced Drex 1.5, a 9B model that returns a probability for every option in one forward pass. It scores 58.08 on Decision Index 0.3.1 and reads up to 128K tokens.
More: shattered.io
Friday 09 October 2026
Microsoft releases Microsoft-Decision-1, a fast decision-scoring model built on Qwen
Microsoft released Microsoft-Decision-1, a decision-scoring model for fast decision-making, built on Alibaba's open-weight Qwen3.5-9B. It is available on Hugging Face and Microsoft Foundry.
MicrosoftOpenOpen weights
Alibaba Qwen releases Qwen-Image-2.1-Turbo, an 8-step 7B image model
Qwen-Image-2.1-Turbo is a text-to-image model that generates images in 8 steps, with 7B parameters, on Hugging Face. An uncensored GGUF version has also been posted, with 426 likes and 2,138 downloads.
More: MarkTechPost
Thursday 08 October 2026
JetBrains releases Mellum2.1, a 12B open MoE model for coding agents
JetBrains released Mellum2.1 under Apache 2.0, a 12B mixture-of-experts thinking model with 2.5B active parameters for coding agents. Reinforcement learning on real repositories raised its SWE-bench Verified score from 2.0 to 47.0.