China's Open-Weight Coding Wave: 4 Models, 18 Days

[Alex]: Welcome back to the Nerd Level Tech AI Cast, where the code never sleeps and neither do we—except, you know, on patch Tuesdays. I’m Alex, your resident explainer of things that make your brain hurt.

[Jamie]: And I’m Jamie, your friendly neighborhood question-asker, here to make sure Alex doesn’t go full “whiteboard in the basement” on us. Today, we’re talking about China’s Open-Weight Coding Wave—four models, 18 days, and enough benchmarks to make your GPU sweat.

[Alex]: You know it’s a wild month when you need a Gantt chart just to keep up with model releases. Between April 7th and April 24th, four Chinese labs dropped open-weight coding models like they were hot potatoes. We’ve got Z.ai’s GLM-5.1, MiniMax’s M2.7, Moonshot’s Kimi K2.6, and DeepSeek’s V4. All open-weight, all targeting agentic engineering workflows, and all—here’s the kicker—priced way below what you’d pay for Western frontier models.

[Jamie]: Okay, so my first question—what exactly is “open-weight”? Is it like open-source, but for neural networks after leg day?

[Alex]: [chuckles] Pretty much. Open-weight means the actual neural network weights—the learned parameters—are released for public use, not just the code. So you can take these models, run them on your own hardware, fine-tune them, or just stare lovingly at the 1-trillion-parameter tensor if that’s your thing.

[Jamie]: I mean, who hasn’t spent a Friday night admiring a weight matrix, right? [PAUSE] But seriously, why is this a big deal?

[Alex]: Because until now, the really high-performing coding models—think OpenAI’s GPT-4, Anthropic’s Claude—kept those weights locked up tighter than my high school diary. The “open-weight” wave means anyone can use models that are closing in on frontier performance, but at a fraction of the cost and with more flexibility.

[Jamie]: So, four models in just over two weeks. That’s some serious hustle. Can you walk me through the timeline—and, like, what makes each one special?

[Alex]: Absolutely. Let’s rapid-fire the timeline: - April 7: Z.ai releases GLM-5.1, MIT licensed. - April 12: MiniMax open-sources M2.7, but with a “Modified-MIT” license—more on that drama in a sec. - April 20: Moonshot’s Kimi K2.6 drops, with some spicy architectural choices. - April 24: DeepSeek launches V4—actually two flavors, Pro and Flash, both under MIT. And sandwiched in there, you’ve got Anthropic launching Claude Opus 4.7, and OpenAI dropping GPT-5.5. It was like model bingo for AI nerds.

[Jamie]: Okay, but in terms of “who’s the best,” how do these models stack up? Is there a leaderboard, or is it just a giant flex-off?

[Alex]: Oh, there’s a leaderboard. The SWE-Bench Pro is the big benchmark for agentic coding—think of it as the Olympics for code-generating AIs. As of this wave: - Claude Opus 4.7 leads the publicly available pack at 64.3. - Kimi K2.6 and GPT-5.5 are tied at 58.6. - GLM-5.1 hovers around 58.4 on vendor numbers. - MiniMax M2.7 is at 56.2. - DeepSeek’s V4-Pro is a bit lower at 55.4. And then, like a unicorn, there’s Claude Mythos Preview at 77.8, but that’s invite-only—basically, you need to be on the VIP list.

[Jamie]: So, close, but not quite beating the Western models?

[Alex]: Exactly. The real headline isn’t that open weights overtook the top, but that the “good enough” models just got way cheaper. The price floor dropped out.

[Jamie]: Okay, now you’re speaking my language. How much are we talking? Give me numbers I can cry over.

[Alex]: [laughs] Prepare your wallet. For Claude Opus 4.7, you’d pay $5 per million input tokens, and $25 for output. These new Chinese open-weight models? GLM-5.1 is $1.05 input, $3.50 output. Kimi K2.6 is even lower—about $0.60 input, $2.50 output. MiniMax M2.7 is the bargain bin champ at $0.30 input, $1.20 output. And DeepSeek’s V4-Flash, on promo, is a mind-blowing $0.14 input and $0.28 output per million tokens.

[Jamie]: That’s, like, 1/10th—or even less—of the price for output! So if you’re running a coding agent that loves to talk, suddenly you can afford to let it ramble on and on.

[Alex]: Exactly. For companies building agentic loops—where the model spits out lots of code, plans, or actions—the savings are huge. It’s like going from Tesla pricing to used Prius overnight.

[Jamie]: But what about the tech under the hood? Are these just giant models, or is there real innovation?

[Alex]: Oh, there’s some nerd candy here. Let’s hit a few highlights: - **Mixture of Experts (MoE)**: All four models use this. Instead of one giant brain, think of a council of experts—only a subset “vote” on each token, so you get more power without blowing up compute costs. - **Compressed Attention (DeepSeek V4)**: Makes the model faster and lighter to run—think turbocharging your inference with fewer FLOPs and a smaller memory footprint. - **Muon Optimizer**: DeepSeek switched from AdamW to Muon, which orthogonalizes gradients—basically, it keeps the model’s learning on track at massive scales. Kimi K2.6 also did a Muon variant earlier. - **Agent Swarms (Kimi K2.6)**: You can now fan a task out to up to 300 sub-agents, working in parallel over 4,000 steps. It’s like unleashing a bug-fixing army on your codebase. - **Self-improving Scaffolds (MiniMax M2.7)**: They claim the model ran over 100 rounds of optimizing its own training process. “Model, train thyself.”

[Jamie]: Okay, I love the image of an AI running around optimizing itself like it’s cleaning its own room. But are there any catches with these open weights?

[Alex]: Licensing is the big one. GLM-5.1 and DeepSeek V4 are MIT licensed—super permissive. Kimi K2.6 is “Modified MIT,” which is MIT unless you’re a mega-corp. But MiniMax’s M2.7 restricts commercial use without written permission, which annoyed a lot of folks who thought “Modified MIT” sounded friendlier than it is.

[Jamie]: So, read the fine print before you build your billion-dollar startup on top of M2.7. Got it.

[Alex]: Exactly. And not all benchmarks are apples to apples, because labs use different scaffolds and sometimes cherry-pick tasks. Still, the performance gap is shrinking, and the price gap is a canyon.

[Jamie]: So, bottom line—should we crown a winner? Or is this more of a “horses for courses” situation?

[Alex]: I’d say it’s about fit. If you want bleeding-edge accuracy and can pay for it, Claude Opus 4.7 still leads. But if you want “good enough” coding at a fraction of the price, these open-weight Chinese models are a massive unlock, especially for startups, researchers, or anyone running agentic workflows at scale.

[Jamie]: Plus, you finally get to see what’s inside the black box—no more “just trust us, it’s magic.”

[Alex]: Exactly. And that openness? That’s where real innovation happens.

[Jamie]: All right, I’m officially hyped—and slightly terrified for my cloud bill. [PAUSE] Any last words of wisdom, Alex?

[Alex]: Only this: The frontier is moving fast. Check the licenses, mind the benchmarks, and may your token bill always be under budget.

[Jamie]: This has been another episode of Nerd Level Tech AI Cast! Thanks for tuning in. Smash that subscribe, leave a review, and until next time—keep your weights open and your code cleaner than mine.

[Alex]: Catch you next time, folks! [Outro music fades out]