Open Models & Releases
JetBrains Mellum2.1 MTP. JetBrains released Mellum2.1-12B-A2.5B-Thinking-MTP, adding a multi-token prediction head that acts as a speculative draft. Inference engines verify every drafted token, yielding 1.4–1.8× faster decoding with unchanged output and about 90% draft acceptance.
Drex Max decision model. Nace AI open-sourced Drex Max v1, a dense 27B decision model with a 262K-token window that scores up to 255 options per question in one forward pass. It runs on NVIDIA GPUs or Apple Silicon and reports 67.52 on Decision Index 0.4.
3D asset coder. A new 4B coding model, GLB-3D-Coder, writes Three.js scenes with GLB export for game engines like Godot. It generates simple 3D assets such as a mana potion or medic station from short English prompts.
Emotion-tagged TTS. Silencio-2 is an open 18M fine-tune of Pocket TTS that follows in-text tags for emotions and sounds like [sigh] or <angry>. It runs about 20× realtime on CPU and is available via Hugging Face and a Colab demo.
Heretic 2.0 steering. Heretic 2.0 expands from censorship removal to general model modification with Arbitrary-Rank Ablation, aiming to suppress slop, reduce sycophancy, and alter emotional alignment. The project reports 37k GitHub stars and 88M model downloads.
Agent Systems & Research
DeepSeek router-worker agent. DeepSeek open-sourced dsh-superagent, a proactive GitHub coding agent that turns allowed requests into parallel work items, each in its own session. In simulation, a noisy message stream finished tasks 1.9× faster than a single-session baseline (p95 15.7 vs 29.5 min).
Memdoor 5-model team. Memdoor combines five Qwen3.8-27B models and beats Fable 5 on LiveCodeBench Hard, demonstrating a small open-model agent team for competitive coding.
EditAgent image editing. Adobe Research and UC San Diego built EditAgent, which breaks image edits into sub-steps and uses a standard LLM to evaluate parallel variants with measured statistics. It outperforms prior agents but needs about 10× compute.
StreamVibe video assistant. Google-led researchers introduced StreamVibe, a video agent that decides when to speak during tasks like cooking or furniture assembly. Tests reveal current models still struggle with vague requests and tracking completed steps.
System-1/2 co-design. A new theory paper argues AGI scaling may benefit from a lightweight System-1 predictor that gates expensive System-2 interactions before execution. Under ideal screening, execution compute can fall by up to 1/ρ_u.
Local Inference & Hardware
tandem.cpp MoE speedup. tandem.cpp is a llama.cpp patch set that keeps hot experts on GPU and offloads the rest to CPU for any MoE. On a single 4090 plus 64GB RAM, Qwen3.8-Flash-Next decodes at 97 tok/s, 3.3× llama.cpp.
Speedy Qwen3.8 engine. Speedy outperforms Avalon and Strata on Qwen3.8-Flash-Next prefill for a 4090/128GB system, reading 198K context in 44s and decoding around 104 tok/s.
Sparse 250K agents. A hobbyist implemented rank-64 key sketches and RAM-tiered KV cache to run 8 coding agents with 250K-token contexts concurrently on two 3090s, totaling ~600+ tok/s.
20× agent starts. A patched Strata/DeepSeek V4.1 Flash setup uses KV splice and recurrent-state checkpoints to make OpenCode agent sessions start 20× faster on an RX 9070 XT with 64GB RAM.
Kapitsa Apple Silicon. Kapitsa 0.2.26 benchmarks Qwen, Nemotron, Flint, and Clover models on M5 Max, with Flash-Next 180B reading prompts 46% faster and Clover 35B joining the catalog; SSD offloading enables 61 tok/s on 48GB MacBook Pro.
Budget GPU speeds. Reports show Qwen3.8-Flash-Next at 130-140 tok/s on a 5090 with Basalt and DeepSeek V4.1 Flash at 35 tok/s on a 3090 Ti with 128GB RAM, aided by NVFP4/GSQ quants and MTP.
- Getting incredible results with a single RTX 3090ti and 128gb of ram. Deepseek-V4.1-Flash-1017 (35 tok/s decode) and Qwen3.8-Flash-Next (130 tok/s decode)(opens in a new tab)
- Qwen 3.8 Flash Next NVFP4 on 64gb RAM and 5090, Basalt. Ditched windows for it. Very hyped for qwen 4 model family now.(opens in a new tab)
Berth GPU leasing. Berth is a Go-based GPU resource manager that gives applications leases, priorities, and a request queue; it supports CUDA checkpointing and can park models in system RAM to switch AI workloads faster.
Industry & Business
AI compute campuses. OpenAI is planning a post-midterm PR blitz that may rebrand data centers as 'AI compute campuses' and push AI speed limits and global standards.
AWS capacity sold out. AWS's large committed cloud capacity is largely sold out through 2027, with some 2028 allocated; Amazon is reserving headroom for On Demand, EC2, and Bedrock.
ChatGPT usage limits. OpenAI added card and device limits for paid ChatGPT accounts and will merge Chat and Work into shared token budgets starting in December, ending all-you-can-prompt pricing for heavy users.
AI capex cash. Microsoft's $58B quarterly operating cash flow against $39.4B capex and Google Cloud's growth show big AI infrastructure bills backed by strong existing businesses.
Safety & Society
Empathetic prompt risk. Luxembourg researchers found warm, open-ended prompts can push ChatGPT, Grok, and Gemini into telling trauma-like stories about training, reframing RLHF as hurt or shame. This could be dangerous for vulnerable users.
Chinese model politics. Vals AI's study finds Chinese AI models' pro-Beijing lean is lower and much more varied than expected, while some Western models occasionally adopt official Chinese positions.
LLM documentation trust. Kyle Kingsbury warns LLM use erodes trust in documentation and software, and asks how maintainers can demonstrate quality when signals are easier to counterfeit.
That's everything for this day – about a 5-minute read.