This week in deep learning, we bring you Muse Glimmer 30B: An On-Device Agentic Model and a paper on Good Judgment, Bad Advice: Decision-Model Controllers Raise Agent Cost Without Improving Success.
You may also enjoy FLUX 3 Image Klein, Plan or Retrieve First? A Scale Study of Tool Discovery Architectures, a paper on OpenGCUA: Toward a Fully Open Stack for Generalist Computer Use, 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
Muse Glimmer 30B: An On-Device Agentic Model
Meta publishes Muse Glimmer 30B, an Apache 2.0 agentic model quantized to roughly 4-bit for single-GPU inference, trained on over a hundred languages.
Meta releases Muse Spark Open, an 816-billion-parameter mixture-of-experts model with 42 billion active parameters and a 1M-token context, scoring 39 on the Artificial Analysis Intelligence Index.
Reflection AI open-sources Beam’s weights, post-training recipes
Reflection AI publishes the weights of Beam, a roughly 500-billion-parameter model trained over two months on several thousand GPUs with practically zero cluster downtime.
DeepSeek releases updated V4.1-Flash model it says completes tasks faster
DeepSeek-V4.1-Flash-1017 resolves 77% of DeepSWE v1.1 tasks against 74.2% for the September version, and ships with a Search API returning up to 50 ranked results.
DeepSeek open-sources DeepInfer, the inference framework behind its production services
DeepSeek open-sources the inference framework running one cluster at 155 trillion tokens a day, which it says earns $38 per GPU per hour at off-peak pricing.
FLUX 3 Image Klein: open weights, 5.6B parameters
Black Forest Labs opens a 5.6-billion-parameter image model that gains 91 Elo in text-to-image over FLUX.2 [klein] 9B with 38% fewer parameters.
OpenAI acquires Orchestra Bio to bring AI agents into drug development
OpenAI folds Orchestra Bio’s clinical-trial monitoring agents into its Rosalind life-sciences unit on undisclosed terms, targeting the 10-to-15-year drug approval timeline.
MLOps / LLMOps / AgentOps
Announcing llm-d 1.0: Open distributed inference for agents on Google Cloud TPUs and GPUs
Google ships a Kubernetes-native distributed inference stack built with Red Hat and IBM by more than 500 contributors across 25 organizations, demoed on 1,000 sandboxed agents.
Push-based KV transfer to improve TTFT and TPOT for disaggregated inference in vLLM
A technical blog post about moving KV transfer coordination off the critical path, cutting time-to-first-token roughly 3x and improving time-per-output-token up to 30%.
Four evaluation engineering lessons learned from building agent plugins
A candid blog post about grading execution steps rather than final summaries, after an LLM judge wrongly failed three prompts of an eleven-prompt suite the plugin had passed.
Inside NVIDIA Sentry: Secure AI Agents at Runtime
A technical article about governing agent access from a hardware-isolated BlueField-4 trust domain, correlating network traffic with memory introspection to contain agents within milliseconds.
Beyond Top-K: Building search-confidence guardrails for agentic AI with OpenSearch
A thoughtful article about adding a confidence layer between retrieval and action, scoring magnitude, gap, coherence, freshness and authority to abstain on a graduated ladder.
Vercel introduces durable-harness, extending crash-and-timeout durability to the Pi, OpenCode, Claude Code and Codex harnesses as agent workloads move to the cloud.
Learning
RAG Evals: Change Prompts Without Breaking Your App
A practical blog post about turning failures into eval cases, scoring them with code checks and LLM judges, and tracking cost and quality across a 61-case suite for every prompt or model change in Opik.
Scaling agent teams: reach the frontier at 30% of the cost
A rigorous blog post about organizing coding-agent teams, where a strong lead with 16 cheap planning workers reached 61.8% reward for $155 against an all-strong team’s $510.
Cutting our inference bill by 68% by being obsessive about optimization
An instructive engineering post about cutting an AI code reviewer to $0.77 per review from $2.43, while median review time fell from 4m 09s to 1m 32s.
Plan or Retrieve First? A Scale Study of Tool Discovery Architectures
A research paper about a complexity-dependent crossover where plan-then-retrieve beats retrieve-then-plan on multi-tool queries across ToolBench’s 46,984 API endpoints.
Reducing Reward Hacking by Localizing Bad Behavior Instead of Penalizing It
A research blog post about GRAFT, which confines learning updates from detected reward hacks to parameters deleted after training, preventing even hacks monitors cannot detect.
Screwtape: Adversarial Transcript Editing for Monitor Red-Teaming
A red-teaming blog post about refining attacks inserted into real agent transcripts until monitors miss them, producing 35 evasions from 252 runs, six of which transferred.
Open-weight AI models are 6-7 months behind on cyber
A data analysis about Z.ai’s GLM-5.3 topping the open-weight Cyber ECI at a score the closed-weight frontier reached roughly six to seven months earlier.
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.
cloudflare/security-audit-skill
A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings
Papers & Publications
Good Judgment, Bad Advice: Decision-Model Controllers Raise Agent Cost Without Improving Success
Agent scaffolds increasingly delegate small decisions (whether to stop, verify, or change approach) to cheaper models. We ask whether a typed decision model, Jev 1.13, consulted only at deterministic decision points, makes a Claude Sonnet 5.5 terminal agent cheaper without making it worse. In a pre-specified study on Terminal-Bench 2.1, we hold one agent loop fixed and vary only the controller: none, deterministic rules, a small LLM (Claude Haiku 5.5) answering the same questions, Jev, and a placebo that samples Jev’s development-phase choice frequencies at the same moments. On 21 held-out tasks (315 trials), Jev raised cost per attempt by 19% (geometric-mean ratio 1.19, 95% CI 1.02-1.36) at an identical pass rate (73.0% vs. 73.0%). Jev was cheaper than the small-LLM controller (ratio 0.81) but 22% more expensive than the placebo, and no pass-rate difference between arms was detectable. Offline, Jev ranked finish requests better than the alternatives (AUROC 0.78 vs. 0.62 for rules and 0.60 for the small LLM) but was poorly calibrated: its mean probability that a run was complete was 0.50 when 87% of those runs passed, so it asked the agent to verify work that was already correct. No controller ever chose to abandon a run. Informative decision signals did not become useful interventions; we discuss what that would require.
OpenGCUA: Toward a Fully Open Stack for Generalist Computer Use
Computer use has reached a turning point: the focus is expanding from GUI-only interaction to systems that complete tasks on users’ computers through visual and programmatic interfaces. We define a generalist computer-use agent (GCUA) as a system in which the model chooses how to observe, act, and organize its work across visual and programmatic interfaces. Yet frontier implementations, such as GPT-6 Astra’s computer-use system, remain largely closed, limiting systematic study of their design. We introduce Open Generalist CUA (OpenGCUA), an ongoing project to develop and open the full stack for GCUAs with the community. This report presents the first release of OpenGCUA. Code execution is a key feature of our agent design: it lets intermediate results determine subsequent operations and enables models to apply their existing programming capabilities directly to computer interaction. We also propose harness-to-model decision transfer, moving selected observations and task-state decisions from fixed harness rules to the model. On OSWorld-v2, our full configuration improves task scores across all eight evaluated models, including GPT-6 Astra and Claude Opus 5, raising mean partial score by +19.6 points and reducing reported steps per task by an average of 53.0% relative to batched GUI-only execution. We release our implementation and prompts, evaluation trajectories from all eight models, and preliminary versions of a real-device harness and a cross-device SDK. The harness addresses real-device challenges not fully captured by benchmarks, while the SDK provides a common programming interface across computers, mobiles and browsers. Across four real macOS tasks under our real-device harness, selected Qwen3.8-Max runs use 19.4 times as many reported reasoning tokens as GPT-6 Astra, highlighting substantial differences between models in real-device execution efficiency.
