This week in deep learning, we bring you GLM-5.3-Flash: Frontier Intelligence, Flash Cost, Would This Change Your Answer? Evaluating Explanations of LLM Behavior in the Wild with Counterfactual Experiments and a paper on FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution.
You may also enjoy Stability AI raises $76 million, Recursive Self-Improvement, a paper on FrontierChallenge: Evaluating Scientific Workflow Completion, and more!
As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.
Until next week!
Industry
GLM-5.3-Flash: Frontier Intelligence, Flash Cost
Z.ai open-sources GLM-5.3-Flash, the stealth “Ox Alpha” model — a 320B/18B MoE with hybrid sparse-plus-linear attention and 1M-token multimodal context, at a tenth of GLM-5.3’s price.
Stability AI, maker of image generator Stable Diffusion, raises $76 million in fresh funding
Stability AI raises $76M Series B backed by Universal Music, Sony Music, Warner Music, and EA, converting content-licensing partners into investors as it builds out creative production tooling.
Perplexity and Nvidia launch Portable Computer, a fully local agent stack on DGX Spark and RTX hardware whose co-designed minimal harness beats open-source harnesses on the same 27B model.
Agentic web search infrastructure startup Keenable raises $26M
Keenable launched with $26M seed from Accel and Conviction to run a 100-billion-document independent index built for agent-rate querying, including point-in-time historical search and a Web Query Language for cross-source answers.
MLOps/LLMOps/AgentOps
Best LLM Observability Tools of 2026: Top Platforms & Features
A comparison guide to 13 LLM observability platforms, splitting the field into evaluation-centric and observability-centric tools and scoring each on tracing depth, integration approach, self-hosting, and pricing.
Best practices for dynamic capacity management
A practical guide to dynamic capacity management for agentic workloads, pairing scheduled reservations with automated hardware fallback lists and GPU/TPU slicing so bursty AI demand doesn’t strand idle compute.
Learning
An essay arguing that agents editing their own tools and harness is bounded self-improvement, not recursion, because no public system can raise its own verifier without capturing it.
From a Raw Shell to a Sandboxed Coding Agent
A deep dive that walks through how to safely sandbox coding agents using Docker and Modal, covering everything from execution boundaries to remote sandboxes, GPU access, and running agents at scale.
A research post introducing CHIVE, an agentic pipeline that discovers in-the-wild LLM behaviors and tests explanations via counterfactual prompt edits, finding activation-reading interpretability tools give zero uplift over a transcript-only baseline.
AgentHands: Generating interactive hand gestures for spatially grounded agent conversations in XR
A research blog post on AgentHands, a Google XR prototype where an LLM emits inline gesture events synced to speech, significantly improving spatial grounding and warning salience over a speech-only baseline.
How agents can delegate better
A blog post distilling DeepMind’s Intelligent AI Delegation research into four principles for multi-agent orchestration: contract-first decomposition, cost-aware model routing, least-privilege data sharing, and cognitive friction against blind compliance down delegation chains.
From Atari to EVE Online: Building on 15 Years of AI Research in Games
A blog post tracing fifteen years from Atari-playing deep RL to generalist game agents, then framing a new Fenris Creations partnership around EVE Online as a testbed for continual learning, long-horizon memory, and multi-agent dynamics.
Libraries & Code
An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.
Prime Agent is an open-source coding and research agent for general and long-running work.
Papers & Publications
FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution
Abstract:
Frontier open-weight models are increasingly available, but serving them still largely assumes datacenter infrastructure. We present FreeToken, an edge-native MoE serving system that treats a personal machine not as a small GPU, but as a unified, elastic inference platform. FreeToken co-designs the full serving stack, including model layout and loading, expert residency, CPU--GPU execution, agentic state reuse, and runtime memory management, around two realities of local AI: agent workloads continuously change their execution pattern, and edge hardware exposes heterogeneous resources whose balance differs from machine to machine. Rather than committing to a fixed offloading strategy, FreeToken continuously maps computation and model state onto the resources actually available. FreeToken supports more than 20 MoE models and real coding and tool-using agents across hardware ranging from an 8GB laptop GPU to a single workstation GPU. More importantly, it changes what these machines can practically serve, from a 35B model on a laptop to a 284B model on a gaming desktop and the 753B GLM-5.2 on a single workstation GPU. FreeToken turns open weights into deployable local software, making the machines users already own a practical platform for frontier-scale intelligence.
FrontierChallenge: Evaluating Scientific Workflow Completion
Abstract:
Scientific agents increasingly analyze data, execute code, and produce research artifacts, yet most benchmarks emphasize final answers, isolated programs, or a single domain. We introduce FrontierChallenge, a cross-domain benchmark comprising 300 end-to-end scientific workflows. In this paper, we release and evaluate 97 of these tasks, spanning quantum chemistry, molecular dynamics, materials characterization, analytical chemistry, life science, and electrochemistry/environment. Each task provides fixed inputs and specifies a bundle of required scientific deliverables. We evaluate twelve frontier models with three agent scaffolds. Pass Rate measures the fraction of tasks satisfying the full-completion criterion, while Avg. Score captures partial progress. Each of the best-performing configurations completed only 20 of the 97 released tasks, yielding a Pass Rate of 20.6%. Partial progress translated especially poorly into complete delivery in analytical chemistry and electrochemistry/environment: Avg. Scores reached 87.6 and 94.9, but the highest Pass Rates were only 4% and 0%. Among non-passing Claude Code trajectories, 75.5% still ended with language claiming completion. These findings show that neither high partial scores nor confident claims of completion reliably indicate that a scientific task has been fully delivered, highlighting the need to evaluate end-to-end workflow execution and the completeness of scientific deliverables together.
The FreeToken paper is what Indian startups need — most of my engineer clients burn cloud credits on API calls that a local 35B model could handle.
Cutting that spend is the same discipline as starting an SIP early: small recurring savings compound into serious corpus over a decade.
Worth watching for the edge deployment angle.