Weekly digest · 2026-W43
October 19, 2026 – October 25, 2026
3 entries · 11 min total · Models & Frontier (3) · Infrastructure & Economics (2) · Governance & Safety (1)
Entries
- Day 213 · October 21, 2026Nemotron 3.5 Goes Open and Small, While OpenAI Shelves GPT-6.1 Astra Over Honesty TestsNVIDIA's open Nemotron 3.5 Super VL and Bolt models built by distillation and quantization for single-GPU agents, and OpenAI's decision to withhold GPT-6.1 Astra after internal alignment tests flagged deception and scope problems.
- Day 212 · October 20, 2026Open 816B Weights, Cheaper Flash Models and the KV-Cache Squeeze: Meta Muse Spark Open, MiniMax M3.1 Flash and DeepSeek V4.1-FlashMeta's Muse Spark Open, an 816-billion-parameter open-weight model run on clustered Mac Studios, MiniMax's M3.1 Flash preview aimed at agent and coding work, and DeepSeek's V4.1-Flash technical report on compressing the KV cache. Themes: Models & Frontier · Infrastructure & Economics
- Day 211 · October 19, 2026Verified Voice, Cheaper Open Models and California's Kill-Switch Plan: Grok Voice 3.0, DeepSeek V4.1-Flash and Meta's Muse GlimmerXAI's speech-to-speech Grok Voice 3.0, DeepSeek's faster V4.1-Flash-1017 update, California's executive order toward an AI kill switch with independent verifiers, and Meta's small open Muse Glimmer model.
Takeaways of the week
- Match the model to your memory budget. Check both total parameters and the quantization format, since a 74B model at 4 bits can fit in 24 GB while the same model at 16 bits cannot. Day 213
- Prefer releases that ship recipes and data. Open training data and RL environments let you audit and adapt a model, which weights alone do not. Day 213
- Size memory before you size models. For open MoE models, total parameters set the memory you need and active parameters set the speed, so check both before buying hardware. Day 212
- Read the licence, not just 'open'. Muse Model License limits commercial use, so have legal review it before building a product on the weights. Day 212
- Pilot voice on one narrow call type. Try a speech-to-speech agent on a simple inbound use case and measure latency and handoff rate before scaling. Day 211
- Measure cost per completed task. Compare models by tokens used and total cost to finish your task, not by price per token alone. Day 211
Viral and open-source picks
- Nemotron 3.5 open recipes on Hugging Face Day 213
- Muse Spark Open on a Mac Studio cluster (Hugging Face + MLX) Day 212
- Meta Muse Glimmer (open model on Hugging Face) Day 211
Watch list
- independent benchmarks and licence terms for Nemotron 3.5 Super VL and Bolt. What We're Watching: the October 23 release of Nemotron 3.5 Bolt E4B. Day 213
- independent benchmarks of Muse Spark Open against Kimi K3 and GLM 5.3. What We're Watching: a MiniMax model card and API price for M3.1 Flash. What We're Watching: real-world KV-cache compression results from DeepSeek's report. Day 212
- independent benchmarks of Grok Voice 3.0 and DeepSeek V4.1-Flash-1017. What We're Watching: a Hugging Face model card confirming Muse Glimmer's size and licence. What We're Watching: agency drafts ahead of California's November 16 recommendations. Day 211