AI Daily | Model access, physical tests and infrastructure

Baidu releases ERNIE 5.5 Flash Base weights

2026-11-05 · Seven-day watch

Baidu released a base model with 280 billion total parameters and 14 billion active per token. FP8 and BF16 weights use a custom ERNIE license.

It is intended for further training and has not undergone mid-training or post-training. Deployment decisions need to account for licensing and additional training, rather than treating the checkpoint as a finished chat product.

Baidu: ERNIE 5.5 Flash Base

Tencent Hy4 is available; weights remain a separate milestone

2026-11-04 · Seven-day watch

Tencent says Hy4 is available in products including TokenHub, WorkBuddy and CodeBuddy, with visual understanding and computer-use capabilities. Its announcement separately promises weights and system cards in the following days.

Teams can evaluate the hosted offering; local deployment still needs downloadable weights and verified terms. The reported performance comparisons are largely vendor-organized evaluations, not independent replications.

Tencent: Hy4 launch

Inkling 2 Small launches hosted access ahead of weights

2026-11-05 · Seven-day watch

Thinking Machines released Inkling 2 Small for inference and fine-tuning on Tinker. It accepts text and images, with 594 billion total parameters and 17 billion active.

The company plans Apache 2.0 weights the following week after external safety testing and red-teaming. Hosted access does not establish downloadable availability, and the model does not accept audio.

Thinking Machines: Inkling 2 Small

ERNIE 5.5 Flash Base · Baidu

FP8 / BF16 · 2026-11-05 · stable

Recent verified release: base weights under a custom license; the post-trained edition is still forthcoming.

Baidu: ERNIE 5.5 Flash Base

Checked 2026-11-08T00:36:52.234375+00:00

Hy4 · Tencent

Hy4 hosted · 2026-11-04 · stable

Recent verified release: hosted access is available; downloadable weights remain a separate promised release.

Tencent: Hy4 launch

Checked 2026-11-08T00:36:52.234721+00:00

Inkling 2 Small · Thinking Machines

Downloadable weights · 2026-11-05 · announced

Recent verified announcement: downloadable weights await external safety testing; Tinker inference and fine-tuning are available.

Thinking Machines: Inkling 2 Small

Checked 2026-11-08T00:36:52.235060+00:00

Claude Code · Anthropic

2.1.319 · 2026-11-07 · stable

Previously covered version: adds background MCP tasks; retained as a verified baseline, not another news story.

Claude Code official releases

Checked 2026-11-08T00:36:53.275716+00:00

NVIDIA explains a tiered memory design for agents

2026-11-06 · Recent development

NVIDIA explains how model weights and KV caches compete for HBM. Dynamo software coordinates GPU memory, CPU memory and CMX storage, moving temporarily inactive context out of scarce GPU memory.

The explanation helps clarify why long-running sessions depend on both memory capacity and interconnect speed. This is an architecture discussion, not a chip launch; actual throughput depends on workloads, cache hits and configuration.

NVIDIA: memory and full-stack co-design

ToyBlocks separates robot planning from physical control

2026-11-06 · Recent development

Ai2 introduced ToyBlocks, a tabletop robotics research environment using simulated demonstrations. In a constrained physical block task, a Molmo2–ManiFlow combination succeeded in 15 of 30 trials.

The Franka tests used blocks seen in training and the easiest scenes, with visible upright characters. Safety interventions counted as failures. The 50% rate is specific to these tasks; new objects and complex workstations need further testing.

Ai2: ToyBlocks robot research environment

Amazon Bedrock adds hosted Kimi K3.1 Pro access

2026-11-07 · Recent development

AWS announced Kimi K3.1 Pro (Highspeed) on Bedrock, with explicit prompt caching and access through cross-Region inference.

This adds a model option within existing AWS governance. Teams still need to check routing, data residency and regional compatibility. The event is a new hosting route, not the underlying model's original release.

AWS: Kimi K3.1 Pro on Bedrock

AWS: official announcements feed

Envision reports an operating 800V DC data-center system

2026-11-02 · Seven-day watch

Envision's November 2 announcement says a solid-state-transformer system is powering live workloads at its Ulanqab Galaxy AI campus, converting 10kV AC to 800V DC and using storage to manage load changes.

The design matters for rack density and conversion losses. The campus's 3GW figure is a full-buildout plan, not operating capacity; the announcement does not provide an independent efficiency measurement.

Envision: 800V DC power deployment

BIS posts an extension of the affiliates-rule suspension

2026-11-06 · Recent development

A BIS public-inspection document extends the suspension of expanded end-user controls for affiliates through January 10, 2027. Publication and effectiveness are scheduled for November 9.

Compliance teams should distinguish a posted document from an effective rule: November 9 is still ahead of this issue's cutoff. Extending this suspension does not remove other export controls.

BIS: affiliates-rule suspension extension

Dropbox agrees to buy Tigris for about $350 million cash

2026-11-05 · Seven-day watch

Dropbox announced the planned Tigris acquisition on November 5. Its filing says a definitive agreement was signed November 3 for approximately $350 million cash, subject to customary price adjustments.

Tigris's object storage supports Dropbox's move into AI infrastructure. Closing is expected in the fourth quarter, subject to conditions being satisfied or waived; additional filesystem and content-processing capabilities remain planned.

Dropbox: planned Tigris acquisition

Dropbox: Q3 2026 filing