Reflection AI promises a 501B open-weights model, and GLM 5.3 lands on Bedrock
A big American open-weights MoE is coming, GLM 5.3 lands on Bedrock, and two reminders that the plumbing under agents (disks, CLIs, memory) deserves as much thought as the model.
Two of today’s items are about what an agent runs on, not what it thinks with: a CLI, a sandbox, a shared disk. The model news is there too, and it’s mostly about who gets to host what.
1 · Top story
A US lab is about to ship a 501B open-weights model
Reflection AI announced Beam, a sparse mixture-of-experts model with 501B total and 23B active parameters, trained on 23.8T tokens. Weights are promised under Apache 2.0 this month, with a 256K context during RL and an “effective” 1M after midtraining. The post claims SWE Bench Pro v2-Hard 77.2, Terminal Bench v2.1 80.1 and 3–4× less inference compute than GLM-5.2 on hard reasoning tasks.
Those are the lab’s own numbers, and the weights aren’t out yet. If you’ve been avoiding Chinese-origin open models for procurement reasons, this is the first credible candidate for that slot, but I’d wait for independent evals on your own coding and tool-use traces before planning around it.
2 · Models & features
GLM 5.3 shows up on Bedrock
AWS made Z.ai’s GLM 5.3 generally available on Bedrock on October 5. It’s a 753B MoE with roughly 40B active per token, a 1M token input window, up to 128K output, always-on reasoning with adjustable effort, and explicit prompt-cache points. Access is through US and Global cross-Region inference profiles.
For a team already on Bedrock, that’s a large open-weights model behind your existing IAM, logging and guardrails, with no GPUs to run. The announcement doesn’t list pricing or which APIs are supported, so check both before you wire it into an agent loop.
3 · Agents
Cloudflare rewrites its CLI with agents as the main user
Cloudflare introduced cf, an open-source TypeScript CLI generated from its OpenAPI schemas, covering more than 3,000 API operations against Wrangler’s roughly 280. JSON is the default output, config moves to cloudflare.config.ts, and Wrangler gets 18 months of maintenance. It’s in open beta; this is InfoQ’s report of the announcement.
Schema-generated, discoverable commands are a decent pattern for any platform that expects agents to operate it. The flip side is a much bigger surface to scope: before an agent gets a token for something with 3,000 operations, decide which subset it’s allowed to call.
4 · Research
Fixing agent memory costs more than re-reading, at least on short records
A new arXiv paper asks what an agent should do when a source document changes: patch its stored memory or just re-read. On short records, localized repair used 5–10× fewer revision tokens than rebuilding memory, but every memory method cost at least twice a full re-read. Memory only got cheaper after 2–14 uses on ~10,000-token documents, and source-filtered re-reading was the cheapest overall.
It’s one author, two 7B models and ICU medication records, so treat it as a prompt to measure, not a rule. Still, if your agents cache derived facts about documents that change, it’s worth asking whether the cache pays for itself before you build invalidation logic.
5 · People worth following
Anthropic moves Cowork’s VM and inference into the cloud
Simon Willison quotes Anthropic engineer Felix Rieseberg: the new Cowork runs model inference and the VM in the cloud, and “each session gets its own sandbox, not sharing state with other sessions.” Previously the VM ran locally, which cost users battery and disk.
Per-session sandboxes are the sensible default for agents that run code, and it’s the same direction as the next item. What you give up is local file access, so you’ll want a clear story for how data gets in and out.
6 · Architecture
Cloudflare Containers could expose a previous tenant’s disk blocks
Cloudflare’s write-up, published September 24 and picked up by InfoQ this week, describes a dm-thin pool configured with skip_block_zeroing. A paid-account customer writing 4 KiB into a 64 KiB block could read the up-to-60 KiB remainder left by an earlier container. The researcher found residual data on 18 of 24 production placements; Cloudflare fixed the config, retired running container disks and says it found no evidence of exploitation or any needed customer action.
The microVM boundary was fine here. The leak was in the storage allocator underneath it. If you run multi-tenant agent sandboxes, check what your block layer does on reallocation, because a zeroing flag turned off for speed is easy to miss.
Nothing here needs action today. Bookmark the disk-zeroing one for your next sandbox design review.
Middle East, India & events · 1
- TII released Falcon-Emirati-7B, a model tuned for Emirati dialect — It builds on Falcon-H1-Arabic (hybrid SSM-transformer) and was trained on web text, MSA heritage material and glossary-constrained synthetic dialect data. TII reports 84.83% on its Alyah benchmark; that’s the developers’ own test, so check it before relying on it. Source
Researched and drafted with AI, checked against the linked sources. How we use AI →