AWS opens Bedrock Managed Agents for OpenAI.
Amazon Bedrock Managed Agents (OpenAI)
Preview regions: US East (N. Virginia), US West (Oregon), US East (Ohio). During preview AWS says there is no additional charge for BMA beyond the underlying AWS resources agents consume; pricing may change at GA. Product page frames the Codex harness + AgentCore stack for long-running, permissioned work.
Why it matters after DevDay: Issue 27 covered GPT-6.1 Sol and the Recap’s agent APIs — this is the AWS-side primary for running OpenAI-shaped agents under existing IAM/governance without inventing a second control plane. Do not re-lead Sonnet 5.5 or Sol as today’s lead. Preview ≠ GA.
Sources: AWS What’s New · BMA product · DevDay Recap (context)
None this window.
| Model / package | Org | Notes |
|---|---|---|
| Bedrock Managed Agents Sep 29 · OpenAI-powered |
AWS / OpenAI | Preview 3 regions · IAM/MCP/AgentCore · no extra BMA fee in preview |
| llama.cpp tip b11279 Sep 30 · ~6:58 AM ET |
ggml | Tip GLM-5.3-Flash · GA still v0.5.0 |
| Ollama v0.35.1-rc0 Sep 29 · ~4:14 PM ET |
Ollama | Pre 10 web searches · MLX + llama.cpp bumps |
| PEFT 0.21.1 Sep 29 · after 0.21.0 Sep 15 |
Hugging Face | GA TP fix · Transformers ≥5.17 |
| HyperPod Inference Gateway Sep 24 |
AWS | New vLLM/SGLang · vendor TTFT % claims |
| Astra Ultrafast API docs · service_tier |
OpenAI | Active gpt-6-astra · Sol Ultrafast not GA |
| MLX / mlx-vlm unchanged overnight |
Apple ML / Blaizzy | Watch still 0.32.3 / 0.7.4 from Issue 27 |
| Item | Beat | Status |
|---|---|---|
| BMA IAM + CloudTrail Sep 29 preview |
Per-agent IAM · approvals · audit inside AWS | Watch |
| Google zero-trust agents Pt 2 Sep 15 |
Model Armor · Semantic Governance · Anomaly Detection | Watch |
| OpenAI safety cases / Sol addendum Sep 28–29 |
Guidelines + Critical cyber — not re-led | Watch |
| Sonnet 5.5 cyber / CVP Sep 28 |
Carry from Issue 27 — not re-led | Watch |
Critical · Preview vs GA · tip vs pin
BMA is preview in three regions — design for IAM/MCP smoke, not a production SLA. llama.cpp tip b11279 ≠ GA v0.5.0. Ollama 0.35.1-rc0 stays off the prod pin until non-RC.
Watch · vendor TTFT % · Ultrafast SKUs
HyperPod Gateway latency percentages are AWS-reported. Astra Ultrafast is documented; Sol Ultrafast is still not a clean GA claim — verify the model id against your account.
Opportunity · GLM tip + PEFT TP
Smoke GLM-5.3-Flash on tip b11279 if you already run llama.cpp tips. If you train with PEFT + TP, bump to 0.21.1 with Transformers ≥5.17 before chasing new adapter methods.
Smoke llama.cpp tip b11279 with a GLM-5.3-Flash GGUF
Grab tip b11279 (or newer same-day tip). Load a GLM-5.3-Flash / GLM5-Next GGUF and confirm hybrid-index path runs on your Metal or CUDA box. Keep GA v0.5.0 as the rollback pin. Do not treat tip as production.
Preview BMA or keep Ollama Decision Models on a sandbox
If you have AWS access in us-east-1 / us-west-2 / us-east-2, open the BMA preview APIs and attach one MCP tool under a dedicated IAM role — confirm CloudTrail events. Or stay local: sandbox Ollama 0.35.1-rc0 after 0.35.0 systemone, still with prod on 0.34.4.
Issue 28 · Wednesday, September 30, 2026 · Not investment advice · Sources cited inline · Other news: none this window · Issue 27 Sonnet 5.5 / GPT-6.1 Sol / DevDay / Ollama 0.35.0 / MLX 0.32.3 / mlx-vlm 0.7.4 / safety cases / Grok Bedrock / Jev / Transformers GGUF not re-led · Issue 26 OpenShell / Completions / pause / b11228 / Holo4 not re-led · Issue 25 Ollama 0.40.0-rc0 / mlx-vlm 0.7.3 / Live Avatar / AU Medicare / TRL 1.14.0 not re-led · Issue 23 Genie One MCP / Opus 5.5 not re-led · Do not invent Haiku 5.5 GA · Sol Ultrafast not GA · Listen omitted (episode not published)
Past 1–2 days. ASTGL / Fraser takes — sourced, not recap theater.
BMA — OpenAI agents under AWS IAM
DevDay sold Sol and the Agents API; AWS’s BMA preview is the perimeter story: same OpenAI-shaped harness, but sessions, tools/MCP, and approvals sit inside your AWS identity and audit trail. If your blocker was “agents can’t leave the account boundary,” this is the preview to smoke — not a GA SLA, and not a reason to re-price Sol overnight.
Sources: AWS WN · BMA
llama.cpp tip lands GLM-5.3-Flash
Tip b11279 (~6:58 AM ET Wed) adds GLM-5.3-Flash (GLM5-Next) with hybrid-index memory / k-pool work — MTP stripped from the initial PR. Tip, not GA: pin stays v0.5.0 until you choose a tip build. Worth a smoke if you already chase Zhipu / GLM GGUFs locally.
Source: b11279
Ollama 0.35.1-rc0 — Decision Models keep moving
After yesterday’s 0.35.0 Decision Models drop, v0.35.1-rc0 (Sep 29 ~4:14 PM ET) bumps MLX + llama.cpp (tip b11232), allows ten web searches per response, and adds explicit model capabilities on create. Still pre-release — keep a GA pin on 0.34.4 for production.
Source: v0.35.1-rc0
PEFT 0.21 — more ways to adapt, then a TP patch
PEFT 0.21.0 (Sep 15) shipped KaSA, Super-Tuning, ShadowPEFT, and a Riemannian-preconditioned LoRA optimizer; 0.21.1 (Sep 29) fixes Tensor Parallel with Transformers ≥5.17. If you fine-tune at scale, the Monday move is the TP patch — treat new methods as experiments until your own evals say otherwise.
Sources: PEFT 0.21.0 · 0.21.1
Ultrafast for Astra — Sol still not “GA fast”
OpenAI’s Ultrafast docs put
service_tier: "ultrafast"on GPT-6 Astra for API users (low default TPM tiers; WebSockets recommended). Sol Ultrafast remains a preview / “coming days” story from DevDay — do not bill it as GA. Docs also mention preview wording for “GPT-5.6 Sol”; treat naming as docs-ambiguous until OpenAI clarifies againstgpt-6.1-sol.Sources: Ultrafast mode · GPT-6.1 Sol