Meta releases an 816B open-weight model, and Temporal workers scale to zero
Meta opens the weights of a very large multimodal MoE, Temporal workers scale to zero on Cloud Run, and VPC Service Controls gets folder-level perimeters for agents.
Three of today’s items are really about boundaries: where weights run, where workers live, and what a new agent project is allowed to touch. Less glamorous than a model launch, and more likely to decide whether your pilot survives security review.
1 · Top story
Meta puts an 816B open-weight model on Hugging Face
Meta’s Muse Spark Open went up on Hugging Face on October 15 as a mixture-of-experts with about 816B parameters, 42B active per token, 256 experts and a 1M-token context. It takes text, images, audio and video in and returns text. Builds are on Hugging Face, and AWS lists it on SageMaker JumpStart.
The practical catch is hardware: the BF16 checkpoint is around 1.5 TB, and the NVFP4 build needs Blackwell GPUs. Reasoning can’t be switched off, so budget for latency. The launch post gives no benchmark scores and no license, which is the part to check before anyone builds on it.
2 · Models & features
IBM packs chat, safety and speech into one Granite endpoint
Granite Switch 4.1 (8B and 30B) bundles a base Granite model with a decision adapter, a Guardian safety adapter and a speech-to-text component behind one endpoint. The adapters are activated LoRAs, so a decision can reuse the KV cache from the generation before it, and IBM says VRAM stays close to the base model alone.
That’s a neat answer to the “three small services around every LLM call” problem. The speed and quality claims are IBM’s own, the licence isn’t stated in the post, and the adapter SDK is promised for later.
Try this: add it as one more arm in a routing or guardrail test. Baseline is your current separate classifier call; the metric is end-to-end latency at equal accuracy.
3 · Agents
Temporal workers on Cloud Run can scale to zero
Google and Temporal put Temporal Serverless Workers on Cloud Run in preview on October 15. Temporal Cloud watches backlog-to-worker latency per task queue and adds or removes Cloud Run worker-pool instances, with no cluster or autoscaler to run. Pools can hold GPUs and list at roughly 40% less than instance-billed Cloud Run.
For agent workloads that burst, each model or tool call is a durable activity, so a lost worker means a retry rather than a lost run. Two things to plan for: the first task after idle waits for a cold start, and it requires Worker Versioning, so you roll out new builds rather than patching in place.
Try this: run a worker with min instances at zero and again at one, and compare p95 start latency after ten idle minutes.
4 · Research
Train a small model to break code, then use the breakage as SWE tasks
SWE-at-Scale is a preprint that reverses real bug fixes from 688 repositories to train a small bug-inducing model (a Qwen3.5-9B LoRA, one GPU). It then generates executable tasks, keeping a task only if a previously passing test now fails. Its two-step variant produced usable tasks 94.0% of the time, and humans preferred its issue text to real SWE-bench Verified issues 61.2% of the time.
If you build agent evals or training data from your own repos, this is a cheap way to get tasks that match your code. The authors have released the code and models, linked from the arXiv page, so you can try the recipe on your own repos.
5 · People worth following
Five ways Deep Agents keeps the context window small
LangChain’s Deep Agents team wrote up how its open-source harness fights context rot: tool results over about 20,000 tokens are written to files, with only the head, tail and path left in context. History is summarised at 85% of the window and saved to disk, skills load on demand, noisy work goes to subagents, and memory is a plain file you own.
The file-based memory is the interesting design choice, because it survives a model switch. It’s a vendor post with no accuracy numbers, and the thresholds are defaults you’ll want to tune on your own traces.
6 · Architecture
VPC Service Controls now draws perimeters around folders
Google made folder-level VPC-SC perimeters generally available on October 15, at no extra cost. Projects under a folder inherit the policy, ingress and egress rules can reference folders or whole organisations, and a violation analyser proposes least-privilege fixes you can test in dry-run mode.
Google frames it around agents as non-human identities. The architectural gain is that a new agent project is inside the boundary on day one, instead of waiting for someone to remember. Go through dry-run first: a folder perimeter applied over a messy hierarchy will block things you forgot you relied on.
Friday’s reading ends here. Whatever you skipped will still be old news on Monday.
Middle East, India & events · 1
- NVIDIA’s Nemotron-pretraining-Indic-v1, about 2.8T tokens across the 22 scheduled Indian languages — NVIDIA says continued pretraining on Indic text lifted five of six benchmarks, by 4.98 to 9.82 points per Indic benchmark. The translation pipeline is open on GitHub; the dataset is “releasing soon” and its licence isn’t stated. Source
Researched and drafted with AI, checked against the linked sources. How we use AI →