Excellent LFM3 GGUFs from AtomicChat, as usual. These results also show how tricky it is to quantize small models while preserving BF16 fidelity. I usually aim for 98 or 99 EAR, but that looks tough for this model. 96 EAR might be a more realistic target. I’ll double-check the accuracy on a few task benchmarks. (Also interesting: the difference in next-token probability distributions between FP16 and BF16.)

Today I'll release: - much smaller statistically lossless checkpoints of Qwen3.8 27B, with EAR passing 98, for vLLM (should work also with SGLang but I didn't test). => around 19 GB while still being better than the official FP8 - the smallest LFM3 1.7B GGUF with an EAR passing 98 (this was super hard!! very sensitive to quantization this one) I'm also working on a post-SLQ refinement step, using autoround. It already works but it's quite costly. I will probably release these v2 this weekend. GSQ would also be a good refinement steps for 2-bit and 3-bit layers. I'm checking the cost. I'm almost ready for Qwen4 🫡

TensorFold 1.2.4 is out. 🚀 Two of the measured wins in this release: • DeepSeek-V4.1 restored a saved 319K-token conversation from NVMe in 1.2 to 1.4 seconds, versus 155 to 178 seconds, on two GB10s at CUDA TP=2 with replay prefill. • Flash Next on M5 Ultra: two overlapping clients finished in 47 seconds instead of 127 in the MLX 4-bit parallel-serving test. Also new: Gemma 4 E4B on CUDA, LFM3-1.7B text on Metal and Clef-Flash decisions on HIP. The HIP runtime now builds and loads on Windows too. There are cold-loading improvements, tool-call fixes and clearer API errors. One upgrade note: existing disk-cached prompt states need a fresh prefill after updating. Thank you to everyone behind the code, testing and bug reports. The community puts an incredible amount of work into these releases. @arminxnova @atasrb @Bizuayeu @BlivionIaG @codengod @defilan @devopsfortunato @unigilby @jayleaton @jishvenugopal @Wi11y010 @teleogen @lasse_ai @petruspennanen @dubsthedev @woolsey And on GitHub: AbrahamCete, AdrianBinDC, BobClawblaw, ElderOrb, gordonhatusupy, goshi, jschmied, laidick, ppardus, shuuul and tolewis. @petruspennanen @plotarmordev @volatilemarkts @MiaAI_lab

TensorFold 1.2.3 is out. 🚀 Some useful Flash Next improvements in this one: • NVFP4 on two GB10s at TP=2: plain decode 21% to 23% faster, with 2.62 GiB less model memory per rank. • M3 Ultra: a 1,024-token reply beside a 64K prompt finished in 8 seconds instead of 48 in the parallel-serving test. Same tokens. Those are specific test results, with the configurations and comparisons in the release notes. We've also added PDF document input, Ornith-1.5-122B on Metal, native dots.mocr OCR and Qwen embeddings/reranking on HIP, and LFM3-1.7B on Intel Iris Xe. There are fixes for prompt caching, memory use, streamed tool calls and serving too. A huge thank you to everyone contributing code, testing builds and reporting problems. There's so much work behind each release. @arminxnova @Bizuayeu @BlivionIaG @codengod @devopsfortunato @diwakergupta @unigilby @jayleaton @jishvenugopal @JRB_AI @joshuaswarren @teleogen @lasse_ai @MiaAI_lab @woolsey @petruspennanen And on GitHub: AbrahamCete, BadAd84, BobClawblaw, dinadur, garvrao80, joernchen, jschmied, ppardus, shuuul and tolewis. @MiaAI_lab @volatilemarkts @petruspennanen @plotarmordev

oMLX 0.7.3 is out! Ask AI, the question bar for your local models in any app, now talks back. Press Option-Shift-Space, speak, and your model answers out loud with captions. It can also search the web, read links and look at your screen. Ask AI voice and tools - Voice runs on LFM3-Omni, LFM3-Audio and Qwen3-Omni. With LFM3-Omni 2B on an M3 Ultra, the first audio comes about 100 ms after you stop talking. - Cut in while it's speaking, or share your screen in the middle of a conversation. - Web search, link reading and screen capture work in both text and voice. - Prompt placeholders like {date}, {app} and {clipboard}. Faster DeepSeek V4.1 - INT8 Activation Prefill on M5 Ultra: 64K prefill 2,882 -> 3,507 tok/s (+22%), 195K 2,476 -> 3,200 tok/s (+29%). - 12 concurrent requests decode 56.3 -> 125.7 tok/s (+123%). - Engram prefetch adds another +15% to prefill at 128K. Many requests at once, much less memory - 8 Lightning MTP streams at 81K on Qwen3.8-Flash-Next: peak memory 248 -> 161 GiB (-35%), decode 180.7 -> 241.1 tok/s (+33%). - Batched decode on oQ models: 16 requests 146.9 -> 221.9 tok/s (+51%) on M3 Ultra, 8 requests +74% on M2 Max. - Hybrid Qwen models reuse a shared system prompt across conversations, TTFT 0.89 -> 0.25 s. Also in 0.7.3 - Update downloaded models by fetching only the files that changed. - Better oQ quants, KL drops 9-49% for under 5% more size. - LFM3 text, audio and d1 models, and Unlimited-OCR-2. - Temporary chats in the web UI. Thanks to everyone who sent PRs and tested the rc!

From more efficient computer-use models and open weights to multi-model agent platforms and frontier safety, the period since our last edition brought major developments across the AI-agent stack. Here’s everything worth catching up on in the latest edition of AI Agents Pulse: 1️⃣ OpenAI released GPT-6.1 Terra for visual reasoning, computer use and writing in ChatGPT Work, Codex and the API. Standard API pricing is $1 per million input tokens and $5 per million output tokens. 2️⃣ Google announced Gemini 4 Carbon for coding, cybersecurity and long-horizon tasks. Access is starting with a subset of cyber defenders through the Fairwind Program, ahead of broader developer, business and consumer availability. 3️⃣ Mistral released improved Mistral Large 4 weights for agentic coding and defensive cybersecurity, with Low, Medium and High reasoning modes. Its modified MIT license requires a separate commercial license for businesses operating a Model-as-a-Service offering. 4️⃣ Liquid AI released the first LFM3 models in the 2B-parameter range, bringing text, vision, audio, omni and decision models onto a shared backbone. Base and post-trained weights are available on Hugging Face, alongside Playground and API access. 5️⃣ Google announced Genie 4, a world model for real-time interactive simulation with native audio. Access began with limited trusted testers; Project Genie access for Google AI Ultra subscribers and Gemini API early access are still coming. 6️⃣ GitHub made Agentic Workflows generally available. Developers define automation in Markdown, which compiles to GitHub Actions YAML, using Copilot or another coding agent. Agents run with read-only permissions by default, with proposed changes checked before application. 7️⃣ Microsoft introduced the Copilot harness in Foundry Agent Service in public preview. It manages the agent loop and sandbox for long-running tasks, while developers supply instructions, skills, tools and model choice. 8️⃣ GitHub introduced the Copilot app and Copilot Fusion at Universe. The app connects agents and tools across projects, while Fusion chooses how to execute a task across one or more models. 9️⃣ OpenAI released GPT-6 Rosalind for multistep life-sciences research with scientific tool use. Access is limited to organizations approved for Rosalind Research and their authorized users, rather than general API availability. 🔟 OpenAI updated its Model Spec with guidance on context, oversight and human agency, organized around five values. The company said its models do not yet fully reflect the intended behavior described in the Spec. 1️⃣1️⃣ OpenAI detailed expanded monitoring for frontier agents across training, evaluation and internal use, with adaptations for external deployment. Its layered system combines activation probes, model reviewers, investigative agents and human response to flag unauthorized or deceptive behavior.