Meituan released the weights for LongCat-2.5, a 1.6-trillion-parameter agent model trained entirely on Chinese hardware. The mixture-of-experts model activates roughly 48 billion parameters per token and targets long, autonomous tasks, particularly coding. Its API handles up to one million tokens and generates responses up to 128,000 tokens.
Meituan's self-hosting guide uses 194 Huawei Ascend accelerators: 64 for prompt processing, 128 for generation, and two for images. The company reports 88.4 on Terminal Bench v2.1, versus 89.1 for Claude Opus 5; Claude Opus 5.5 and GPT-6.1 Sol are absent. The MIT-licensed release includes no Nvidia GPU guidance.
Why it matters: For companies evaluating Huawei hardware, Meituan's 194-chip recipe offers a concrete deployment path for a 1.6-trillion-parameter model, while Nvidia users get no deployment instructions from the release.
Jensen Huang’s resignation could shrink Nvidia’s unusually generous flow of capital to AI startups. Nvidia participated in 105 funding rounds this year, often as the largest investor, backing neoclouds, inference providers, and open-weight model companies.
Investors expect interim CEO Colette Kress to be more selective and write smaller checks. Under Huang, infrastructure companies seeking $100 million to $500 million for GPUs often raised roughly half from Nvidia at valuations of $1 billion or more, giving other investors confidence to follow.
The pullback would hit CoreWeave, Crusoe, Lambda, Mistral AI, and Black Forest Labs, while traditional VCs may pause deals or mark down valuations. Nvidia’s board is searching for a permanent successor.
Why it matters: For traditional VCs, the underwriting shortcut is gone: without Nvidia supplying roughly half of many $100 million-to-$500 million GPU raises, billion-dollar valuations face a harder sell.
Amazon Bedrock now traces AI inference spending to individual IAM principals without requiring API changes. The capability covers Claude Code, Claude Desktop, OpenAI Codex, ChatGPT Work, and OpenCode when they use Bedrock’s bedrock-runtime or bedrock-mantle endpoints.
With IAM Identity Center and federated sign-ins, each call carries the user’s identity into AWS Cost and Usage Reports (CUR 2.0), where session tags can add team, cost-center, and project dimensions. Billing aggregates by principal and usage type hourly or daily; per-prompt detail requires invocation logs or request metadata.
Previously, users sharing a fixed IAM role and session name could collapse an organization’s spending into one principal. AWS also supports workload attribution through Application Inference Profiles, Projects, and Workspaces.
Why it matters: For teams choosing between Bedrock and Anthropic’s managed offerings, the operational split is clear: Bedrock puts per-user dollars in CUR and Cost Explorer, while Anthropic keeps usage visibility and native spend limits in its admin systems.
Weight-only quantization can halve raw model-weight storage by replacing FP16 or BF16 values with INT8. That matters during autoregressive decoding, where moving weights for every new token can take longer than the computation itself at small batch sizes. Activations remain in FP16 or BF16, preserving the standard floating-point input path.
PyTorch’s TorchAO supports post-training configurations including INT8 and INT4, with per-channel or per-group parameters that can better preserve quality than one scale for an entire tensor. GPTQ and AWQ go further by using calibration activations to reduce output error.
The trade-off is deployment complexity: packed layouts, group sizes, zero-point conventions, kernels, and hardware must match. GGUF is a container, not an algorithm, and measured speedups remain hardware-dependent.
Why it matters: For inference teams running small batches, reducing weight movement can address a memory-traffic bottleneck without quantizing activations. But TorchAO, GPTQ, AWQ, and GGUF deployments still require exact agreement between checkpoint format, kernels, model architecture, and target hardware.
Anthropic says outdated prompting habits are limiting Claude Opus 5.5. The company advises users to drop phrases such as “think hard” and “think carefully,” because the model reasons automatically and adjusts depth to the selected effort level.
Anthropic’s internal tests found Opus 5.5 at Medium matched predecessor Opus 5 at High, while removing the old phrases produced faster responses with no measurable quality loss. Higher settings cost more and do not consistently improve results.
The company also recommends giving the model a complete brief in one message, attaching charts as files or images, and requesting finished documents rather than drafts. Teams comparing Opus with cheaper Sonnet should account for the extra steps and effort settings.
Why it matters: Anthropic is shifting prompt engineering from coaxing Claude into reasoning toward specifying the end result and letting Opus 5.5 manage the workflow. For teams paying for higher effort on Sonnet, the company’s claim that those costs can approach Opus pricing makes model selection worth revisiting.
GS Caltex now deploys machine-learning models in a few hours, down from about a week. The South Korean energy and chemical company automated the handoff between process engineers building models without code in Amazon SageMaker Canvas and ML engineers managing SageMaker Pipelines.
Previously, ML engineers connected Canvas assets manually by setting production data paths, supplying model configuration, and wiring preprocessing to inference jobs. The new workflow reuses validated Data Wrangler preprocessing flows and trained models in production.
GS Caltex runs batch inference every 30 minutes through Amazon EventBridge instead of keeping continuously running endpoints, and says the change meaningfully lowered inference costs. The reported gains remain the company’s claim.
Why it matters: GS Caltex can now move a validated Canvas model into a repeatable production workflow without rebuilding the handoff each time, while its batch setup avoids the need for continuously running endpoints and the company reports lower inference costs.