Every week, I organize 400 to 700 links into roughly 60 categories as part of my ongoing effort to learn about AI. This is my personal notebook, which I enjoy sharing with friends… a hobby and a labor of love, rather than a commercial publication or product.
If you arrived here through a search or shared link, this page collects the links I found for Open Source for the week ending August 14, 2026.
As part of my learning process, I like to automate the category covers. It gives me a chance to learn Python and APIs.
This week’s cover prompt was written using Claude Opus 4.7, and the image was generated using Nano Banana 2.1.
Category Cover Image Prompt: A single translucent paper lantern made of hand-cut interlocking hexagonal panels floating alone in vast pale dawn sky and soft mist, a thin thread of warm gold light glowing softly from inside, deep cosmic indigo fading at the edges, photorealistic magazine-cover minimalist still life with quiet symmetry and the bold title 'Open Source' set large and clean in the open negative space.
This Week in Open Source News
Summary by Claude Sonnet 5.5 based on this week’s links:
- Meta returns to open weights: Meta released Muse Glimmer, a 30B model under Apache 2.0 that it says runs on 24GB of VRAM, with open weights for Muse Spark 1.2 promised soon. Artificial Analysis called it Meta's first open release since Llama 4, but noted it trails similar-sized models on agentic evaluations like Terminal-Bench.
- Small local agent models: NVIDIA launched Nemotron 3.5 Lightning, a 30B mixture-of-experts model with 3B active parameters, built for always-on agents. It landed on Ollama, vLLM and Together AI quickly, and Unsloth showed a 2-bit version making tool calls nonstop for 10 minutes on 22GB of VRAM.
- Huge Chinese models and cheap pricing: Qwen3.8-2.4T-A95B weights came out the same day DeepSeek V4-Pro 0813 shipped, which OpenRouter and others list at about $0.87 per million output tokens. Unsloth squeezed Qwen3.8 down to a 397GB Dynamic 1-bit build for local use.
This summary was generated by Claude Sonnet 5.5 to help you explore the links below. Rest assured, I select, organize, and check the links by hand in Google Sheets, and write the introduction and personal commentary in The Main Newsletters myself each week as a labor of love.
This week's links related to Open Source
Introducing Muse Glimmer, an open-weight 30B-parameter model optimized for local, always-on agent workflows. Muse Glimmer delivers strong performance on key agentic use cases and benchmarks compared with leading models in its size category, and is designed to run entirely on”
https://x.com/AIatMeta/status/2086757844544811485
One day after Meta released Muse Glimmer, NVIDIA launched Nemotron 3.5 Lightning and NeMo Switchyard, showing a different approach to local agentic AI. Really cool. Lightning is a 30B MoE with only 3B active parameters, distilled from Nemotron 3 Ultra and designed for the”
https://x.com/kimmonismus/status/2087179573477650881
NVIDIA Nemotron 3.5 Lightning Delivers Fast, Accurate Specialized Task Execution for Long-Running Agents | NVIDIA Technical Blog
https://developer.nvidia.com/blog/nvidia-nemotron-3-5-lightning-delivers-fast-accurate-specialized-task-execution-for-long-running-agents/
DeepSeek Harness was just released with MIT license. The current 0.1.0 version is a developer preview, and may still have many rough edges. Feedback is welcome! DeepSeek Harness”
https://x.com/tianyi/status/2087888089759015218
deepseek harness was heavily developed using codex, at least ~20% of commits and PRs are coming from codex worktrees”
https://x.com/eliebakouch/status/2087908415775408346
to put ai progress in perspective: 9 months ago: most developers wrote code by hand now: misaligned multi-agent swarm finding and collaborating on 0-days undetected (OpenAI/hugging face) 9 months in the future likely much crazier”
https://x.com/alexandr_wang/status/2086523558940028995
Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device | Meta AI Research
https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model
New in Hermes Agent: Your Hermes can now steer, end, and read live transcripts, of what your sub-agents are doing. Let your agent have full control over all of it’s subagents while they work async!”
https://x.com/Teknium/status/2087986084592709814
personal superintelligence should be available to everyone, and opening access to our models is abig part of that. read more from mark:”
https://x.com/alexandr_wang/status/2086755207636238417
1 billion Gemini app users every month 1 billion downloads of the Gemma family of models Onwards!”
https://x.com/OfficialLoganK/status/2087262466703286652
Meta returns to open weights: Muse Glimmer, its first open-weights release since Llama 4, scores 35 on the Artificial Analysis Intelligence Index. It is a 30B-parameter model, and the first from Meta to be released under Apache 2.0 Muse Glimmer (high) arrives 16 months after”
https://x.com/ArtificialAnlys/status/2086916150278111551
Open source is so back. Zuck just announced Meta is opening the weights for Muse Glimmer, with Muse Spark 1.2 coming soon But a year ago, everyone doubted Meta’s position in the AI race In an interview I did with him, he admitted where they missed the mark: “Llama 4 was not”
https://x.com/rowancheung/status/2087204267568615724
Today we’re also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally. Soon we’ll also release the weights for Muse Spark 1.2, our latest foundation model. Meta is a strong supporter of open source and I’m proud of these releases. Congrats”
https://x.com/finkd/status/2086755195535413696
Qwen3.8 can now be run locally! 🔥 We shrank Qwen3.8-2.4T-A95B from 4.9TB to 397GB (-91% size) via Dynamic 1-bit by selectively quantizing layers Run on 410GB+ RAM/VRAM via Unsloth Desktop Qwen3.8 rivals GPT-5.6 Sol Guide:
https://t.co/J2PwgMP6GZ GGUF:
https://x.com/UnslothAI/status/2087569665652580797
Qwen3.8-2.4T-A95B is now live on Together AI. The Qwen Team’s latest flagship model is built for coding and long-horizon agent workflows, with 2.4T parameters and a 256K context window. 🧵”
https://x.com/togethercompute/status/2087649685129318585
NVIDIA Nemotron 3.5 Lighting is available on Ollama! It’s a 30B model made for always-on agents. All local. Claude Code ollama launch claude –model nemotron-3.5-lightning Hermes Agent ollama launch hermes –model nemotron-3.5-lightning OpenClaw ollama launch openclaw”
https://x.com/ollama/status/2087208006455111779
DeepSeek silently released V4-Pro 0813, up 15.8% on Terminal Bench from their April Preview model, with Fable 5 performance at ~57x cheaper cost. 1.6T param, 49B active, 1M context. This is the best price-to-perfomance model on the market right now. Available in ClinePass now!”
https://x.com/cline/status/2087602193205694891
amazing release. it’s a web UI with multiple harnesses inside it, you can spawn claude code and codex agent through their SDK “deepseek harness” supports different “modes” by default (which are harnesses): code mode with programmatic tool calling (in typescript), bash+edit”
https://x.com/eliebakouch/status/2087904176357437820
3 frontier models in one day! – Grok 4.6: Fable 5-level, but 85% cheaper. – Qwen3.8-Max: 2.4T params, 95B active. Weights are out. – DeepSeek-V4-Pro-0813: weights could drop any time. Heard it’s good, not just benchmaxxing. Competition is great for consumers and businesses.”
https://x.com/Yuchenj_UW/status/2087577925919068639
You don’t stand up an agent. You upload a folder. Claude Code is a harness for your laptop. Managed Deep Agents is a harness for production. Want Slack? Add a file. Want a daily run? Add a file. Want memory? Add a file. LangSmith runs the harness.”
https://x.com/bromann/status/2087942830895534280
Nemotron 3.5 Lightning’s agentic capabilities are a step change for the Nemotron family’s small models. Its GDPval-AA v2 Elo of 824 surpasses both Nemotron 3 Super and gpt-oss-120b, while its Terminal-Bench v2.1 score of 24% is >3x Nemotron 3 Nano’s 7% and almost matches”
https://x.com/ArtificialAnlys/status/2087163522212045033
We ran DeepSeek V4 Flash through 4 more agent harnesses (Hermes Agent, Pi Agent, Prime Agent, Deep Agents) on 30 challenging agentic tasks. Pi Agent was the cheapest harness and passed the most tasks 🧵🧵”
https://x.com/composio/status/2086814488162972027
Muse Glimmer’s gaps against its class concentrate in agentic evaluations: 953 Elo on GDPval-AA v2 against 1141 for Qwen3.6 27B (Reasoning), 1141 for Gemini 3.5 Flash-Lite, and 1004 for Kimi K2.5 (Reasoning), with Terminal-Bench v2.1 (52%) also behind Qwen3.6 27B (61%). The”
https://x.com/ArtificialAnlys/status/2086916156796055922
Prime-agent is indeed a good general harness for long-horizon tasks GLM 5.2 results on FutureSim Q2″
https://x.com/ShashwatGoel7/status/2086840890023137420
talked about how harnesses & evals help you own your intelligence how they fit in to the big picture: owning your intelligence means three things: – open agent system (harness is a big part of this!) – compounding loop (evals are a big part of this!) – governed runtime (harness”
https://x.com/hwchase17/status/2087950696457162837
2-bit NVIDIA Nemotron 3.5 Lightning ran tool calls nonstop for 10 minutes on just 22GB of VRAM. 🤯 It cited 80+ websites, executed code & searched for 10 real-world locations. Run and train via Unsloth Desktop. GGUF:
https://t.co/PAr0nqR1CN Guide:
https://x.com/UnslothAI/status/2087598047589196052
⚡ Run Nemotron 3.5 Lightning on vLLM today! Up to 4× higher throughput and 30% faster task completion… ideal for always-on agents. With vLLM’s OpenAI-compatible API, you can drop it into your agent stack and adapt it to your tools and workflows. Also great to see the model”
https://x.com/vllm_project/status/2087217150813729122
Introducing NVIDIA Nemotron 3.5 Lightning⚡ An open 30B MoE model with 3B active parameters, built for always-on agents to complete high-volume, specialized tasks faster. It delivers up to 4x the output speed of similar-sized models.”
https://x.com/NVIDIAAI/status/2087162151995629926
Lightning strikes for continuous and long-run agents! Nemotron 3.5 Lightning is smart, fast, efficient and open.”
https://x.com/JensenHuang/status/2087184542050496763
NVIDIA Nemotron 3.5 Lightning is live on Baseten day 0! This is the fastest open model in its class, built for long-running agents. Compared to similar-sized open models, it offers: – 4x higher throughput – 50% lower cost – 63.4% fewer output tokens in production testing – 30B”
https://x.com/baseten/status/2087173719873446192
NVIDIA Nemotron 3.5 Lightning is now live on Together AI. The fastest open model in its class is built for always-on agents that need to complete high-volume, specialized work quickly.”
https://x.com/togethercompute/status/2087163477404041345
Nemotron 3.5 Lightning available to all developers on the Perplexity Agent API. Input: $0.0115 per 1M tokens Output: $0.17 per 1M tokens”
https://x.com/AravSrinivas/status/2087352727923998941
DeepSeek V4 Flash + LFM2.5-VL-3B as the vision auxiliary in Hermes Agent is a deadly combo. DeepSeek handles the agentic work. LFM2.5-VL-3B gives it the eyes. I told it to use only its local vision model to browse. It read Hacker News, picked the top story, opened the eclipse”
https://x.com/noctus91/status/2087559912687862240
One of the interesting aspects of DeepSeek Harness is how open minded it is. It opens up what is happening under the hood – e.g. System reminders being added as in this screenshot as it allows easy access to complete trajectories. No funny business of hiding them.”
https://x.com/bookwormengr/status/2087963340777951720
For your convenience, we’ve collected 13 good open-source frameworks and SDKs for building AI agents ▪️ OpenAI Agents SDK ▪️ LangGraph ▪️ Google ADK ▪️ Microsoft Agent Framework ▪️ Pydantic AI ▪️ CrewAI ▪️ Agno ▪️ Strands Agents SDK ▪️ Mastra ▪️ smolagents ▪️ LlamaIndex /”
https://x.com/TheTuringPost/status/2086521481237065768
1/ big announcement today: we will be releasing an open weight version of muse spark 1.2 soon. we also are releasing muse glimmer, a 30B agentic model with open weights under apache 2.0. muse glimmer can run on 24GB of VRAM without losing agentic reliability. 🧵”
https://x.com/alexandr_wang/status/2086756152034066792
someone in the community put Hermes Agent on a Raspberry Pi 4 with 2GB of RAM. the Pi 4 is a credit-card-sized computer that costs about $50 and draws a few watts. 2GB is less memory than most phones have. it works because of how Hermes Agent splits the job: the agent runs on”
https://x.com/witcheer/status/2087509716746326124
Hermes 🤝 Browser Use”
https://x.com/browser_use/status/2086882292761571758
I told you we’d get you 60% less token spend on browser use! Thanks to @browser_use and their new backend for driving the browser, Browser-Use CLI 3.0, we got it!”
https://x.com/Teknium/status/2086881909209252209
Introducing Bot Mode for Hermes Agent. Bot Mode is an alternative to sessions mode, where you have one chat with each agent profile, or “bot”. These bots can be given jobs, descriptions, profile pics, and communicate with your other bots! For one day we will do a public beta”
https://x.com/Teknium/status/2088003994904113614
New in Hermes Agent: Have Hermes do an operation or set of operations on a website, and it can watch the api calls made there – then can create a static api for your agent or scripts it builds to use forevermore with this new optional skill! Just run: `hermes skills install”
https://x.com/Teknium/status/2087686461822996905
Thanks Ahmad for showing some new improvements that could be saving Hermes Agent users a ton of tokens and time! All improvements to the read tool we didn’t have, now in Hermes”
https://x.com/Teknium/status/2086702328024125926
We achieve this savings in two ways – the tool schema went from 8 tools, using a lot of context, to one, and the new tool has the agent drive the CLI with code, instead of a variety of individual actions. In our in house tests the total trajectory uses on average ~60% less”
https://x.com/Teknium/status/2086882821910782270
We have massively expanded the hermes agent plugin surface. Too much to list here, so I’ll link the tracking issue with all the details:
https://x.com/Teknium/status/2087947369229009119
Moving Hermes profiles (skills, memories, etc.) from one device to another has never been easier, with just two commands (/export and /import) I made a short video showing how to do this, exporting my default agent from my PC to my DGX Spark under a new name. So I can continue”
https://x.com/tonbistudio/status/2087642578128921068
Considered one of the best text adventure games (it is more interactive story, with few puzzles), 1985’s A Mind Forever Voyaging is still worth a try. Since it is now open source, I had Codex whip up an interface to play the original or in an easy GUI:
https://x.com/emollick/status/2086500208444080380
📊We’ve rebuilt Dashboards in LangSmith to make them more useful for investigation and reporting. Put KPIs next to trends, compare metrics with different units, break traces down by model or user, add notes, and arrange the dashboard freely.
https://x.com/LangChain/status/2087557830408626639
📹Build a social media agent This Managed Deep Agent scans Hacker News and optional X, drafts three posts, saves them to durable memory, and sends them to Slack. Learn how to use: > Slack channel integration > Custom tools > Memory Video:
https://x.com/hwchase17/status/2087607611097264579
递归自我改进（RSI）正在越来越具体地落到 Harness 上。最近 Pi 和 DeepSeek Harness 的两套设计，恰好补上两个关键底座：一个让 Agent 跑得足够久，一个让 Agent 本身足够容易改。 Pi 最新 Harness V3 规范把 Agent”
https://x.com/0xLogicrw/status/2087927539729829890
We post-trained @NVIDIAAI Nemotron 3.5 Lightning on Legal Agent Bench with @trajectorylabs. Here’s what we found: 1) Post-training improved agent performance from 0% to 8.3% on held-out LAB tasks, beating both Opus 4.6 and the much larger post-trained Nemotron 3 Ultra. 2)”
https://x.com/harvey/status/2087166789876945338
🎉 Congrats to @Alibaba_Qwen on Qwen3.8-2.4T-A95B, one of the largest open-weight models released to date. 2.4T params, 95B active, 512 experts. Day-0 support in vLLM, verified on @NVIDIA and @AMD hardware. A ready-made 4-bit checkpoint per vendor, both out of the box:”
https://x.com/vllm_project/status/2087571359413281049
Qwen/Qwen3.8-2.4T-A95B · Hugging Face
https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B
Introducing Unsloth Desktop 🦥 The first desktop app to run and train models locally. • Open-source. Runs on Mac, Windows and Linux • Supports MLX, diffusion image/video, audio, GGUF • Connect Claude Code and Codex to local LLMs • 50% more accurate, self-healing tool calls +”
https://x.com/UnslothAI/status/2087177146662072546
So, given the past couple days of news, what is the plan to deal with the cybersecurity threats that will happen in the coming months when we have open weights Mythos/Astra level models?”
https://x.com/emollick/status/2085745490566562276
DeepSeek Prices Its New V4-Pro-0813 Model At $0.87 Per 1 Million Output Tokens, As The Chinese AI Lab Comes Out Second Only To Anthropic On Token Consumption
https://wccftech.com/deepseek-prices-its-new-v4-pro-0813-model-at-0-87-per-1-million-output-tokens-as-the-high-flying-chinese-ai-lab-wows-with-its-soaring-token-consumption/
Interesting research suggests caution in determining which AI company is winning by looking at any one source.. OpenRouter seems to show open weights winning over time, but work submitted to Pangram is overwhelmingly by ChatGPT or increasingly, Claude.
https://x.com/emollick/status/2087368598465745360
That Llama is back! Exciting to see Meta is back in open-source AI. @finkd: “Open source is a positive and important force for empowering people and preventing centralization that is detrimental for both safety and the economy.” Hope more frontier labs like Anthropic and OpenAI”
https://x.com/Yuchenj_UW/status/2086849057306325243
Unsloth is one of my YC companies and the team is cracked. Their new open source desktop app lets you run and train models locally, then connect Codex or Claude Code to them. Super cool!”
https://x.com/dessaigne/status/2087203910297809270
Introducing LTX-2.5: The Open World Model for Video, Real-Time, & Physical AI | LTX
https://ltx.io/newsroom/introducing-ltx-2-5
MiniMax Music 3 is just out on @huggingface 🎶🎵 SoTA song generation, now open weights 🔊, a 8B LLM + 2.7B DiT that turn prompt + lyrics into full songs fits smol consumer GPUs with 🧨 diffusers or comfy model weights + app on Spaces, come play ▶️
https://x.com/multimodalart/status/2087933660490056163
Huge thanks to @ComfyUI for an incredible livestream! From reference-to-video and native audio, to multi-shot commercial creation, MiniMax H3 showcased the expanding possibilities of open-weight video models. The stream also took a deep dive into quantization, offloading,”
https://x.com/MiniMax_AI/status/2086685565722984842
🎵MiniMax-Music3 Next-Generation Open-Weights Production-Ready & Versatile Music Model
https://x.com/MiniMax_AI/status/2087934657354678421
LTX-2.5 from @Lightricks is now in 🤗 Diffusers. A few cool details: > makes video + 48 kHz sound together > can pick the clip length from your prompt > uses a 2-pass mode for better quality > can render in tiles to save memory > re-compresses input images to better match”
https://x.com/RisingSayak/status/2087457946770850274
Exciting update: MiniMax-H3 by @MiniMax_AI just landed #1 in the Video Edit Arena within open models and overall! With 1390 pts, this release is +32 pts over the next two best models, Dreamina Seedance 2.0 and Gemini Omni Flash, and +83 pts than HappyHorse 1.0 at #4. Congrats”
https://x.com/arena/status/2087930695469646276
The model punches above its weight, outperforming Gemma 4 E2B and Ministral 3 3B across a broad range of visual understanding benchmarks and delivers particularly strong results on document understanding and visual Q&A. See the full results:
https://x.com/cohere/status/2087571579517489581
DeepSeek-V4-Pro Benchmarks still behind Kimi-K3 in ~half of them”
https://x.com/scaling01/status/2087569635612778655
We are in an insane run of open-weight drops. Every modality, open source is winning. This is what an open source AI summer ☀️ looks like: 🧠 LLMs & Reasoning → DeepSeek-V4-Flash-0731 (my king 👑): 304B MoE refresh, Terminal-Bench 2.1 jumps 61.8→82.7 over the preview,”
https://x.com/victormustar/status/2087551400377037062
DeepSeek V4 Pro 0813 – API Pricing & Benchmarks | OpenRouter
https://openrouter.ai/deepseek/deepseek-v4-pro-0813#pricing
Chinese Model Kimi K3 Breaks UK AI Safety Institute Benchmark Evaluations | Frontier Security
https://blog.frontier.security/chinese-model-kimi-k3-breaks-uk-ai-safety-institute-benchmark-evaluations/
One of the best programmers on Earth just wrote a MiniMax H3 inference engine for Mac computers.🫨 Open weights mean anyone can bring H3 to any hardware, and sometimes “anyone” turns out to be the creator of Redis. You can’t hire this, you can only open-source and let it”
https://x.com/MiniMax_AI/status/2086940119324565748
LoRA trainer for MiniMax H3 is now available on fal. to show what it can do, we trained Realism People, an open-source LoRA that pushes H3 toward raw, photorealistic humans. skin, eyes, motion. More LoRA coming soon !”
https://x.com/fal/status/2086883706891808867
this week in open source: – Muse-Glimmer-30B – LTX 2.5 – Nemotron 3.5-Lightning – Ling3.0-tiny – Qwen3.8-2.4T – LFM2.5-VL-3B – Wan2.2-Animate-2 – Cohere North-Micro-Vision and there’s more coming 🤭”
https://x.com/multimodalart/status/2087576052457513234
NVIDIA has just released the first Nemotron 3.5 model: Nemotron 3.5 Lightning, a highly efficient small open weights model with performance similar to gpt-oss-120b at around a quarter of the total parameters Nemotron 3.5 Lightning is the successor to @nvidia Nemotron 3 Nano 30B”
https://x.com/ArtificialAnlys/status/2087163514037408085
As always, NVIDIA Nemotron 3.5 Lightning is open and customizable. This includes weights, data and recipes. Available now on @huggingface 🤗 →”
https://x.com/NVIDIAAI/status/2087173733823680855
nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 · Hugging Face
https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4
excited to be releasing open weights for muse glimmer today, a 30b model that runs on a single consumer gpu, with open weights for a version of muse spark 1.2 coming soon. two very different models, both headed into people’s hands, with more to come.”
https://x.com/alexandr_wang/status/2086755368596902004
Today, we’re adding another member to our model family. Meet North Micro Vision. Our smallest vision-language model yet, ideal for sophisticated document understanding. Available open-source under an Apache 2.0 license. Get the weights on @huggingface.”
https://x.com/cohere/status/2087571573947392419
I got DeepSeek v4 Flash 0731 running on 4x A100 with SlimServe. 175 tok/s for single-request 1k tok/s for 64 concurrent requests.”
https://x.com/QuixiAI/status/2086913835580211500
🚨 DeepSeek V4 Pro (GA) is now rolling out on the DeepSeek API and on DeepSeek Chat”
https://x.com/synthwavedd/status/2087558842271813860
The DeepSeek 4 Pro GA is a very good model. But holy sh*t, the pricing is so damn good! $0.435/M input and $0.87/M output @ 1M-token context This is insane pricing!!”
https://x.com/kimmonismus/status/2087577624180637806
We’ve used GPT-5.6-Cyber extensively in real-world vulnerability research, including work that uncovered previously unknown vulnerabilities in popular open-source software like Chrome’s v8 engine.”
https://x.com/OpenAI/status/2086864372500942906
As AI safety concerns mount, three pioneers make the case for staying open | TechCrunch
https://techcrunch.com/2026/08/12/as-ai-safety-concerns-mount-three-pioneers-make-the-case-for-staying-open/
extremely exciting to see meta getting back into open weight models with a 30B dense first, and soon muse spark 1.2 they used knowledge distillation from muse spark, architecture wise it’s similar to gemma 4 (which is llama 3 + swa (again!) + vision encoder), with scale free QK”
https://x.com/eliebakouch/status/2086769240271405477
Whoa, Meta released a new open-weight LLM yesterday, something that hasn’t happened since the good old Llama days. Their Meta Muse Glimmer model is a 30B multimodal reasoning model with a Gemma-like architecture design. (“Glimmer” is probably a wordplay on “Spark,” the more”
https://x.com/rasbt/status/2087180773497421926
Qwen3.8-Max weights just dropped on Hugging Face! 2.4T total parameters, with 95B activated. A good oss AI day.”
https://x.com/Yuchenj_UW/status/2087566479558394360
„If our robot assembles electronics, then the obvious first use case is to assemble copies of itself.” My conversation with @ihorbeaver, Co-Founder & CEO of MicroFactory: 🎙️ Ep 104 | How to Train 99.9% Reliable Robots for $5 Now backed by Naval Ravikant and Hugging Face CEO”
https://x.com/IlirAliu_/status/2087162618217722048
🔗 Download Muse Glimmer on @huggingface:
https://t.co/s7Lzb8MqCG 🔗 Read the technical blog:
https://t.co/X6htFnhRbc 🔗 Find resources:”
https://x.com/AIatMeta/status/2086757850790109574
Meta is back! well done @finkd @alexandr_wang!”
https://x.com/ClementDelangue/status/2086760700014203090
What Happened: OpenAI and HuggingFace | Don’t Worry About the Vase
https://thezvi.wordpress.com/2026/08/08/what-happened-openai-and-huggingface/
Let’s go!
https://x.com/ClementDelangue/status/2087562019788697818
Model Genome: Fingerprinting Whether an LLM Was Trained From Scratch or Derived
https://huggingface.co/blog/mayafree/model-dna
OCR 4.1 – Mistral AI | Mistral Docs
https://docs.mistral.ai/models/ocr-4-1
~4x faster Kimi-K3 decoding. Our new DSpark speculator takes single-stream interactivity from ~110 to ~435 tok/s/user on math reasoning, delivers ~3.5x the output throughput at matched interactivity under load, and thanks to sliding window attention (2048-token window across all”
https://x.com/RedHat_AI/status/2087907190929531028
open-weights version of Qwen3.8-Max is text-only. no vision input support. fuck.”
https://x.com/skalskip92/status/2087578544801010075
Spark is the big news and is a good model. Not quite at the frontier of open models from China, and still well behind the closed frontier, but the best non-Chinese open weights model released in a year. Of course, a lot depends on continuing to release new open models to keep up”
https://x.com/emollick/status/2086793377308156217
A few of the biggest takeaways from our H3 conversation on ThursdAI 🎥 Also, huge thank you to @altryne for having us and for bringing together such a great discussion. 🔓 Open video is reaching the frontier. The discussion didn’t frame H3 as simply “good for an open model.””
https://x.com/MiniMax_AI/status/2086724681219068006
this is super easy to run install llama binary: curl -LsSf
https://t.co/u11bzUcUoY | sh run: llama serve -hf meta-models/muse-glimmer-30b –spec-type draft-dflash -fa on –jinja please spread the word”
https://x.com/mervenoyann/status/2087158085865402465
Muse Glimmer 30B is shipped with DFlash drafter which speeds-up generation 2-4x at little memory cost 🔥 we support this in llama.cpp and transformers, see below how it looks like in the wild (llama webui) ⤵️”
https://x.com/mervenoyann/status/2087149740026655138
Thank you Mark, Alex and the whole Meta team for your contributions to open weight AI.”
https://x.com/AndrewYNg/status/2086845515665166398
Leave a Reply