Muse Ember is Meta’s smallest and most efficient model, now available on the API at $0.10 per million input tokens and $0.50 per million output tokens. It targets repeatable agent tasks such as tool calls, coding subtasks, routing, extraction, and long document parsing. The model offers a 1M-token context window and up to 512K output tokens. It claims more intelligence per dollar than Claude Haiku or GPT Luna.
Key points
Pricing: $0.10 per million input tokens, $0.50 per million output tokens.
Context: 1M-token context window, up to 512K output tokens.
Sources
Control Room halts model workloads within 35 seconds
Control Room halts model output within 35 seconds of an intervene decision. It connects every research inference workload and works with misalignment monitoring systems as part of a holistic safety approach.
Key points
Control Room tracks model workloads across the research infrastructure.
It connects inference workloads, agentic sampling for training and evaluation, data processing, and more.
Sources
dot adds multiple local computer support for tasks
dot now lets you connect multiple local computers to the same dot and pick which one runs a task. It can build and test your app on a dedicated machine while you keep coding on your laptop.
Key points
Connect multiple local computers to the same dot and select the machine for a task.
Build and test your app on a dedicated machine while coding on your laptop.
Sources
Vocal synth models vocal tract as network of tubes
Vocal synth models the vocal tract as a network of branching acoustic tubes, unlike vocaloid which stitches together and pitch-shifts phoneme recordings. The creator is tuning it against recordings of real singers to improve naturalness.
Key points
Model uses a network of branching acoustic tubes to simulate the vocal tract.
Tuning is being done against recordings of real singers to improve naturalness.
Sources
Google TPU 8i Max and TPU 8tx announced in 8th generation expansion
Google expands its 8th generation TPU portfolio with TPU 8i Max, built for massive context windows and highest compute density per rack, and TPU 8tx, purpose-built for secure.
Key points
TPU 8i Max: built for massive context windows and highest compute density per rack
TPU 8tx: purpose-built for secure
Sources
Mythos 5.2 pre-deployment testing by UK AISI
Mythos 5.2 pre-deployment testing was conducted by the UK’s AISecurityInst and results are in its system card. The post highlights that the UK has built something world-class in AISI.
Key points
Pre-deployment testing of Anthropic’s Mythos 5.2 by the UK’s AISecurityInst
Results summarised in Mythos 5.2’s system card
Sources
icwm.cc first major update venue and deadline
icwm.cc announces venue in Manhattan and paper deadline December 22nd. Submission is single-blind, free-form with a page limit and AI allowed. Reviews are public on OpenReview and each reviewer gets fewer than three papers.
Key points
Venue: venue confirmed in the heart of Manhattan
Deadline: paper deadline is December 22nd
Review load: < 3 papers per reviewer
Sources
- Original post
- icwm.cc - Linked in the post
Inkling 2 Small free on OpenRouter limited time
Inkling 2 Small from thinkymachines is free on OpenRouter for a limited time. It pairs vision with agentic coding and can be used at $0 inside Claude Code, Codex and Hermes Agent.
Key points
Free on OpenRouter: limited-time $0 access.
Vision with agentic coding: builds and refines websites, games, 3D designs.
