Open models recap: more on Kimi K3, Qwen 3.8, Xi's WAIC speech, distillation, the open-closed gap, and what's next

What it claims

This is an opinion conversation, not a measured study. Its practical content is first-hand usage reports on Kimi K3 and GLM 5.2, some adoption constraints, and one argument with a stated basis: distillation from frontier APIs matters mainly at the SFT stage, and the open literature has not shown that a stronger teacher yields better SFT data.

Evidence

All of these are anecdotes from the speakers, not benchmarks.

Kimi K3 on real work. Florian used it on Prime Intellect's internal framework. He puts it at roughly the level of the "55, 54" Codex-model versions, below the newest.

• Its code is simpler and more readable.
• It misses niche edge cases that a follow-up Codex pass catches.
• It is usable for supervising runs and running experiments.
• It was the only model that produced a genuinely novel analysis in Nathan's open-ended data-analysis test. It found that Reddit discussion of a model leads its download growth by one to two months.

GLM 5.2. Florian rates it "Sonnet-ish". He picks it for speed: 200–300 tok/s through an API, or a fast internal endpoint. He suggests Kimi K3 as the main agent with GLM 5.2 as the sub-agent for grunt work.

Teacher strength. Open Thoughts 3 and Open Thoughts Agent are cited as the main open reasoning-SFT work. The state-of-the-art open SFT set is still built on QwQ-32B. Regenerating it with stronger models (e.g. GLM 5.2) has not reliably helped, and Nathan says it is unknown whether Claude or Gemini traces would improve an OLMo model.

Cost of using frontier APIs as RL graders. Thinking Machines' final RL run reportedly used about 20–40M rollouts. Grading at that volume through the strongest frontier APIs would be prohibitively expensive and slow.

How to use it

Kimi K3

• Weights: not released at recording. They were expected July 27.
• Licence: unannounced. Moonshot committed to open-sourcing it but not to specific terms.
• Hardware: one full node of B300s just to load the weights. Fine-tuning is expected to need substantial engineering first. Inference-provider optimisation will lag the faster rollouts seen for roughly 500–700B MoEs.
• Subscription: plans at $40, $100 and $200 per month. The $200 tier gives 1M context and appears to get API priority. Other users report constant API errors.
• Latency: servers are in China, so wall-clock time is much higher than GPT.

Other models mentioned as adoption candidates

• DeepSeek V4 Flash: the "workhorse" to watch. V4 Pro is not expected to be a breakthrough. A preview endpoint is being updated near-daily.
• Thinking Machines Inkling: about 1T parameters with decent, not frontier, scores, designed to be fine-tuned with Tinker. A model about a quarter of that size ("Inkling small preview") is expected within weeks and is pitched for domain-specific automation.
• Tencent Hunyuan: about 250B, now under Apache. Earlier Tencent licences barred use in the EU, UK and South Korea.
• Gemma: matches small Qwen models on adoption, but is less established as a research base.

Limits

• Licences may be restrictive. MiniMax has made its licences progressively more restrictive. Whether it reverses course after Xi's speech is open. Check terms for any of these models before adoption.
• Possible US restrictions. DC is discussing restricting Chinese open-weight models in the US, possibly as a "shadow ban" via legal threat. Factor this in for enterprise use.
• Data claims are uncertain. The hosts put "error bars" on what Chinese labs told them about their data practices. External data purchasing is described as rising but unquantified.

Predictions (theirs, unverified):

• Frontier tier: Kimi and Zhipu. Close competitors: DeepSeek and Qwen, with MiniMax likely to join via a 1T+ model.
• A US lab (Nvidia, Thinking Machines or Reflection) may join the top cluster by late November.
• No open model over about 3T total parameters this year.
• Qwen's upcoming open flagship is expected to be weaker than its small models, relative to rivals.