From mid-July to early August — under three weeks — I counted five things that happened around Chinese open-source models. Three are real, two are still rumors.
On July 16, Moonshot AI shipped Kimi K3. My last post covered the process of wiring it in — usage tracking and prefix cost, not whether it was actually good to use. On July 19, Alibaba's Qwen team put out a Qwen 3.8 Max preview, widened to general availability on August 3. On July 31, DeepSeek pushed the GA release of V4 Flash, versioned DeepSeek-V4-Flash-0731. Those three already happened.
Two more are still rumors, with no official announcement behind either. Zhipu's GLM-5.3: founder Jie Tang posted on X in late June asking what features people wanted in the next GLM, and Zhipu's own docs briefly leaked a reference to "GLM-5.3" in early August before it was pulled. Moonshot's Kimi K3.1: a few outlets report it's "targeting an August launch."
I've been moving too. Kimi K3 is now my daily driver, DeepSeek is my go-to for editing code, and I wired in Qwen 3.8 yesterday and I'm still testing it. Meanwhile my old primary, Claude Code Opus 4.8/5, has been getting on my nerves lately.
This post keeps two things separate: what in this wave is verified versus still rumor, and what actually changed for me after switching models for a few months.
Three real, two still rumors
Here's the timeline, fact-checked.
| Model | Status | Timing | Notes |
|---|---|---|---|
| Kimi K3 | Released | 2026-07-16 announced, 07-27 full weights | 2.8T params, 1M-token context |
| Qwen 3.8 Max | Preview → wider rollout | 07-19 preview, 08-03 wider | Not open-weight yet, API-only |
| DeepSeek V4 Flash GA | Released (GA) | 2026-07-31 | Versioned V4-Flash-0731, MIT license |
| GLM-5.3 | Rumor / leaked | No official date | Zhipu says "still in training" |
| Kimi K3.1 | Rumor | Reportedly targeting August | Not confirmed by Moonshot |
Three things worth spelling out.
First, Qwen 3.8 Max isn't strictly "open source" yet. Alibaba currently only offers it via API; open weights were "promised for next week" as of when I checked, still not shipped, and the license terms aren't public. It's fine to fold it into "a wave of open releases" as a narrative — it's genuinely part of the same three-week burst — but its current openness isn't the same thing as K3's or DeepSeek's, which already have weights out.
Second, GLM-5.3 and Kimi K3.1 both remain rumors. In the Zhipu documentation leak, staff privately told the person who posted screenshots that the model is "still in training," with a stated target of matching Fable-class capability — the closest thing to an official signal, but still not an announcement. Kimi K3.1 is even thinner: a few outlets cite unnamed "insiders," and neither Moonshot's blog, GitHub, nor X account has mentioned the name. I'm treating both strictly as rumor, not confirmed fact.
Third, DeepSeek's V4 Flash GA update is described in one line in the official API docs — the model was updated from preview to DeepSeek-V4-Flash-0731, calling convention unchanged. Same scale and architecture as the April preview; it's a retrain, not a bigger model, aimed mainly at agent capability — tool calling, code editing, multi-step tasks. That direction lines up with my own "editing code got noticeably better" impression.
Is Kimi K3 actually that good
In the two days after launch, the claims going around my group chats — and that I repeated in my last post — were: ranked 4th globally, #1 on the frontend code Arena, higher than Fable 5 on SWE Marathon. This time I checked each one.
"#1 on the frontend code Arena": confirmed. On Arena.ai's (formerly LMArena) Frontend Code Arena, K3 debuted at #1 with 1,679 Elo, ahead of Claude Fable 5's 1,631 and GPT-5.6 Sol xHigh's 1,618, winning 6 of 7 subcategories.
"Higher than Fable 5 on SWE Marathon": confirmed. SWE Marathon measures whether a model can carry a long-horizon software engineering task through to completion. In Moonshot's own evaluation table, K3 scores 42.0 on this benchmark, ahead of Claude Opus 4.8's 40.0, GPT-5.6 Sol's 39.0, and Claude Fable 5's 35.0. I checked the official numbers twice; they're consistent.
"Ranked 4th globally": this one's the interesting case, because it's a moving target. Artificial Analysis (an outfit that specializes in model intelligence-index evaluation) titled its own article "K3 achieves #3 on the Intelligence Index" — that was launch week, K3 at 57, behind Fable 5's 60 and GPT-5.6 Sol's 59. Chinese media reported "4th" at the time; I couldn't pin down exactly why the two numbers diverge, possibly a different counting method. But checking the live leaderboard now, K3 has slipped to #7, because Claude Opus 5 shipped afterward and took several of the top spots. Neither "#3" nor "#7" is wrong — they're just snapshots from different points in time. That's how benchmarks work: true the moment you screenshot them, history two weeks later.
Interestingly, Moonshot's own blog was pretty candid about it — as quoted by SCMP, the wording was that overall performance "still trails the most powerful proprietary models" while "consistently outperforming other tested models." Moonshot never claimed to be #1 itself; the loudest version of that claim came from outside the company.
How the trade press covered it
This wasn't just me imagining a wave — several reputable outlets covered it at the same time.
SCMP's July 20 piece quoted Ryan Fedasiuk of the American Enterprise Institute saying the US-China AI capability gap "has now closed to something approaching a matter of weeks," and that K3's performance "toppled the conventional wisdom that Chinese models lag six to eight months behind American ones." TechCrunch's launch-day coverage noted that the market expected K3 to "perform at par with or even surpass Anthropic's Opus 4.8," while Moonshot closed a funding round valuing it at $31.5 billion around the same time.
A second SCMP piece on August 3 grouped Kimi K3 and Qwen 3.8 Max together, calling it "a second DeepSeek moment — but this time, a one-two punch." I think that framing is accurate: last year's DeepSeek moment was a single strike; this time it's Moonshot, Alibaba, and DeepSeek landing hits within three weeks of each other — a different order of density.
Bloomberg ran an opinion piece titled "Anthropic and OpenAI Face a New Threat from Cheap Chinese AI," arguing that if you define the race by developer adoption rather than headline benchmarks, China is closing fast, as more American startups turn to cheaper Chinese models to cut costs. I couldn't pull the full text to verify the exact wording — I can confirm the headline and the gist, but I'm holding back on the specific phrasing.
What actually changed for me
Everything above is verified public information. What follows is my own experience over the past few months — to be clear, this is a feel, not a benchmark I ran myself.
Kimi K3 is my primary model now. It gets work done without a lot of fuss, and it's steady. This is separate from the setup problems I complained about in my last post — that was about usage tracking not reporting correctly in my own tool and prefix costs I hadn't managed, engineering problems on my end, not the model's performance. Once those were fixed, a few weeks of daily use later, K3's feel is: give it clear instructions and it just does the work, no detours.
DeepSeek is cheap, and its code-editing results surprised me. Checking the pricing this time is when I actually saw how cheap — V4 Flash GA output is $0.28/M, versus Kimi K3's $15/M, roughly a 50x gap. DeepSeek's official direction for this GA update is training toward agent capability — code editing, multi-step tasks — which matches my own experience.
Qwen 3.8 just went in yesterday, still testing, no verdict yet. Alibaba widened access on August 3; I wired it in the next day. Not enough hours on it to say anything useful yet — I'll write about it separately once I've tested more.
Claude Code Opus 4.8/5 has been underwhelming lately. A few specific things:
It talks a lot, often circling for a while before getting to the point, and I end up having to pick the actual answer out of a pile of preamble myself.
Give it an answer after asking a question, and it doesn't necessarily follow it. One example from my own experience: I once asked it to place a project somewhere, told it explicitly to create a new project, and it still put the thing inside the project I was currently working in — that's one specific case I ran into, not a claim that it happens every time.
It drops things along the way. This is my own subjective comparison: if I ask it to get to a ten, it tops out around a six.
Add those up and that's why I've stopped defaulting to it as my primary. But I haven't canceled — if I'm honest about the reason, it's that Fable 5 is still there. Fable is Anthropic's strongest tier right now, and that's what's keeping the subscription for me, rather than switching entirely elsewhere. This isn't "Claude doesn't work" — it's that the Opus tier's recent performance, for me, doesn't earn the word "primary" anymore.
Wrapping up
Fact-checking this wave, the most surprising thing wasn't which model scored highest — it was how little a "ranked #X globally" claim holds up over time. K3 was #3 the week it launched, #7 two weeks later, and neither number is a lie — they're just snapshots from different moments. That's what benchmarks are: true the instant you screenshot them, a historical photo shortly after.
What didn't change is the feel I built up from actually using these models for months: K3 gets to the point without the noise, DeepSeek is cheap and genuinely capable, Qwen 3.8 still needs more time, and the Opus tier really isn't what it used to be.
Sources: Moonshot AI's official blog, Kimi K3 GitHub evaluation, Artificial Analysis, SCMP: Kimi K3 launch, SCMP: Silicon Valley reaction, SCMP: Qwen 3.8 Max, DeepSeek official pricing docs, Caixin: DeepSeek V4 Flash GA, TechCrunch.