Hey readers! The model that quietly topped OpenRouter's coding charts under a mystery codename finally has a name, a license, and a story that involves zero Nvidia hardware. Grab a coffee, because this week the open-weights coding world got noisy.

LongCat-2.0: China's 1.6T Open-Weights Coding Model - Meituan uploaded the full weights for its 1.6-trillion-parameter open coding model to Hugging Face, after a June 30 "weights coming soon" placeholder finally went live. - Creative AI News

Here's the fun part: this was the model developers had already been using under the codename "Owl Alpha," topping coding leaderboards before anyone knew who built it.

The design is built for agentic work: a Mixture-of-Experts setup that activates roughly 48B parameters per token, a native 1-million-token context window, and pricing on OpenRouter of $0.75 per million input tokens and $2.95 per million output. Under an MIT license, you can fine-tune, redistribute, and build products on it, which matters far more for your procurement conversations than a fractional benchmark lead.

The detail worth pausing on is the training claim. Per Meituan, LongCat-2.0 was trained and served end-to-end on Huawei Ascend chips using Huawei's HCCL library over more than 30 trillion tokens, with no Nvidia hardware in the loop.

Treat that as a vendor account rather than an independently verified fact, but if it holds up, the supply-chain implications are bigger than the leaderboard delta. A frontier-class coding model that doesn't touch Nvidia silicon changes the math on who can afford to train one.

LongCat isn't arriving in a vacuum. A cluster of open models dropped in the same window, and each stakes out a different corner.

Ornith-1.0: The Open Model That Writes Its Own Agent - DeepReinforce's MIT-licensed family teaches the model to generate and refine its own agent scaffolding during reinforcement learning, rather than bolting code generation onto a fixed human-designed harness. - Emergent

The 397B flagship reportedly scores 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified, surpassing Claude Opus 4.7 on both. If self-scaffolding pans out, it undercuts a lot of the custom harness plumbing teams maintain today.

Tencent's Apache-licensed Hy3 takes on GLM-5.2 at half the size - Tencent's Hunyuan team shipped the full 295B-parameter Hy3 (21B active) under Apache 2.0, reversing an earlier preview license that had excluded the EU, UK, and South Korea and stalled some enterprise evaluations. - VentureBeat

The honest headline is in the title: Hy3 leads GLM-5.2 in agentic search, tool orchestration, and long-context retrieval, but GLM-5.2 still wins on coding benchmarks. Tencent's own numbers report the hallucination rate dropping from 12.5% to 5.4% between preview and full release. Kilo separately moved Hy3 to General Availability, free for a limited time across its VS Code sidebar, cloud agents, and CLI.

Z.ai launches ZCode to challenge Cursor, Claude Code and GitHub Copilot - The Beijing lab formerly known as Zhipu shipped a free desktop "Agentic Development Environment" for its GLM-5.2 model, aimed squarely at long-horizon project work. - VentureBeat

GLM-5.2 is described as a 744B MoE with 40B active parameters, a one-million-token context window, and MIT-licensed weights on Hugging Face. Pricing runs from $16.20/month (Lite) to $144/month (Max), which NEWSx notes undercuts Cursor's $200/month top tier. The Decoder points to a Snowflake comparison across 103 tasks where GLM-5.2 and Opus 4.7 nearly tied after three attempts. The recurring theme across all these launches is cost and openness, not raw peak performance.

GitHub Copilot Just Added Its First Open-Weight Model - Copilot added Moonshot AI's Kimi K2.7 Code as a selectable model, its first break from a picker of closed, US-hosted options. - Singularity.Kiwi

It's rolling out to Pro, Pro+, and Max first, with Business and Enterprise off by default pending admin review. GitHub still hosts the weights on Microsoft Azure, so "open" here is about auditability and cost tiers, not local execution. Worth flagging: GitHub's own docs warn the model "may be less aligned than other Copilot models," per Let's Data Science.

A few other Copilot and tooling updates landed in the same stretch:

Agentic AI Code Review Trust Gap - A Qodo/Gatepoint survey found 94% of engineering leaders already use AI coding tools, but production reliability keeps stalling adoption. - Futurum Group

More capable, cheaper models don't fix the review bottleneck. That's the throughline connecting Sonar Vortex, which claims to cut issues produced by 92% by verifying inside the agent loop, and Tessl Agent, which mines your PRs and session logs to turn recurring mistakes into reusable skills.

That's the quiet lesson under all this week's model news. Whether your agent runs on LongCat, Kimi, or GLM, the harness and the review discipline around it are still doing most of the heavy lifting. Pick the model that fits your budget and license, then spend the saved money on making its output trustworthy.

See you next week.