Jonathon's AI Wiki

Chinese Open-Weight Models — A Vendor-Neutral Decision Framework (Nate B Jones)

9 min read

Source: raw/US_AI_Dominance_Is_Over_-_Here_s_Why.md

Creator: Nate B Jones | Platform: YouTube | Feed: AI News & Strategy Daily | Nate B Jones | URL: https://www.youtube.com/watch?v=JBzz53HqMEs

Against the “China has caught up” narrative around Kimi K3 and Qwen 3.8, Jones argues that “Chinese models” is a broken shorthand — the labs share a country of origin, not a price, deployment path, hardware burden, or best use. His answer up front: yes, serious AI users should test Chinese models, but you must make three separate decisions — the task, the model, and the deployment path — and measure cost per accepted result rather than token price. Bias note: the video plugs his own Ringer multi-agent product and his Substack deep-dive, so the swarm recommendations double as marketing.

Key Takeaways

  • The models are wildly different products. DeepSeek pursues extreme inference economics (15/M output) and called it open weight while the weights were not yet downloadable at recording. Z.ai’s GLM 5.2 is MIT-licensed and very strong at long-horizon coding, but its released BF16 checkpoint is ~1.5TB. MiniMax M3 is 427B total / 23B active per token, spanning coding, computer use, and image/video understanding, under a custom license requiring attribution and notice, prior written authorization above $20M annual revenue, and prohibiting military use.
  • The US government’s own read (CAISI, May): DeepSeek V4 Pro is “the most capable Chinese model” it tested, “about eight months behind” the leading US frontier — and it ranged from “53% cheaper to 41% more expensive per correctly solved task across seven benchmarks.” Token price and finished-work cost can point in opposite directions; DeepSeek has changed prices since, so treat the numbers as an indicator of where to test, not a verdict.
  • Cost per accepted result is the gold standard: include input, output, reasoning, tool calls, retries, latency, and infrastructure. A cheap model gets expensive via long reasoning traces, unnecessary tool calls, late failures, and human cleanup; a premium model like Fable 5 can be cheaper if it succeeds in one pass — and Opus 5 launched with the same token-efficiency focus, which is how frontier models now compete.
  • The distillation fight matters: Anthropic says DeepSeek, Moonshot, and MiniMax used about 24,000 fraudulent accounts to generate more than 16 million Claude exchanges while evading access restrictions; a later White House memo described “industrial-scale extraction by China-based entities” without naming labs or validating Anthropic’s numbers — then Kimi K3 was named by the White House, and K3 was separately publicly accused of distilling Fable 5. These are allegations, not court findings, and the public record doesn’t show how much released capability came from the alleged campaigns.
  • Self-host only when four conditions all hold (see below) — otherwise use a first-party API or managed third-party host.
  • Verdict: use Chinese models selectively — as specialists, challengers, and in some workflows the default — but “country of origin is just the very beginning of having a conversation about due diligence.”

The Three Decisions

1. Name the task (and your failure tolerance)

  • High-volume, price-sensitive API work — extraction, classification, first-pass research, test generation, reviewable work: DeepSeek absolutely belongs in the test set; a 15–30x price cut lets the machine “check and recheck and check again.” Be cautious wherever one ambiguous answer drives an unrecoverable action.
  • Local/personal-hardware work — smaller Qwen models and the R1-distilled models are the real local-AI story (not the same artifact as full R1 or V4 Pro): private notes, offline work, predictable availability, trading capability for control and privacy. The right comparison is not small-Qwen vs Kimi K3 — it’s whether the small model is good enough for the exact job you want to keep on your machine.
  • Long-horizon coding/research/agent work — test GLM, Kimi, Qwen, and MiniMax against the best American frontier models with the same tools and the same acceptance standard. Don’t change your bar: GLM 5.2 is particularly strong on coding, but a model can write an excellent patch and still lose a repository-level task by picking the wrong tool, missing a test, or failing after 20 minutes.
  • Ambiguous/frontier-judgment work — the strongest American frontier systems remain the baseline to beat, with more mature tool integrations, enterprise controls, and operational evidence; Chinese offerings counter on price, open weights, and deployment choice.
  • Sensitive work — decide where the data may go before choosing the model.

2. Name the artifact and deployment

  • “Open” is not one property. Ask separately: Can I call it through an API? Can I download the weights? Does the license permit my use? Can my hardware run it at the context length and speed I need?
  • Mixture-of-experts is why enormous models can still be economical: the router activates a fraction of parameters per token (MiniMax M3: 23B active of 427B), but total parameters still drive storage, memory, networking, and serving burden — a model can be cheap through an optimized API and absurd on your laptop. Don’t convert active-parameter counts into a hardware estimate; ask about checkpoint size, precision, context requirements, and serving topology.
  • The 4-condition self-host test: (1) weights available and license permits your use; (2) the data/sovereignty requirement is real enough to need the model on-prem; (3) the workload justifies the AI infrastructure economics; (4) you have a team that can operate the stack safely. Self-hosting also needs a named owner for authentication, model-server patches, capacity, and observability — otherwise you’re converting vendor risk into an operating problem you discover when something breaks.
  • Three deployment paths: first-party API (provider controls service, price, logging, availability), third-party host (run open weights in your chosen region and contract — an American host running Qwen weights sits under American jurisdiction), self-host (maximum control, total responsibility).

3. Trace the data and the exit path

  • China’s interim rules govern generative-AI services offered to the public inside mainland China — they don’t automatically govern raw weights privately operated outside China. DeepSeek’s first-party privacy policy says personal data is processed and stored in China and inputs may be used to improve its technology; Alibaba says Model Studio customer data is not used for training and offers deployment scopes that can exclude mainland China.
  • Self-hosting removes provider-side runtime filters but not refusal behavior or political bias learned in the weights — you may need to fine-tune, and you must check inference calls, tools, telemetry, and logs. Moving the weights changes who can see the data; it doesn’t make the rest of the software supply chain trustworthy by magic.
  • Own your evals and preserve an exit path: identify who receives the prompt, where it’s stored/processed, governing law and contract, retention and training terms — and whether you can lift prompts, tests, and tools to another model if the provider changes its price.

Distillation and Policy Context

  • Normal distillation is ordinary and disclosed: a teacher generates examples/answers/reasoning traces, a student trains on them. OpenAI sells distillation as a product; frontier labs use it internally (a Mythos teacher model teaches the released versions of Mythos and Fable); DeepSeek openly fine-tuned Qwen-derived students on 800,000 samples curated from R1.
  • A distilled student is not the teacher in a smaller file: it inherits reasoning patterns on examples it saw but loses breadth, reliability, and safeguards — hence “big model smell.” “DeepSeek runs locally” conflates the local student with the frontier teacher.
  • The controversy begins at authorization and conduct: training on API outputs against contractual restrictions, fake accounts, access-control evasion at scale. Chinese entities are typically not authorized to distill American models.
  • Why it matters even if you reserve judgment: chips are physical, model outputs are information. Restricting a teacher’s GPUs doesn’t control what it can teach through an API (his Napster analogy) — capability moves through synthetic data, fine-tuning, RL, and distillation, imperfectly but fast and cheaply enough. That means capability gaps can close faster than hardware policy suggests, and frontier providers tighten access, monitoring, and account controls in response.
  • Export controls didn’t invent Chinese efficiency — MoE, low-precision training, sparse attention, quantization, and hardware utilization reflect a broader decades-long tendency; controls raised the strategic incentive. “The Americans don’t have to worry about chips, they have to worry about power.” China’s AI-plus policy supports open-source sharing of models, tools, and datasets; open weights also reduce dependence on American APIs, though the policy doesn’t state that motive explicitly.

Try It

  • Run the individual-scale version of his protocol: take 20 examples of real work, include the ugly edge cases, pick one Chinese model, run the test, and measure cost per accepted result across the whole chain.
  • His own result doing this: building his Ringer multi-agent swarm (Fable orchestrates, cheaper models execute), Qwen proved a genuinely useful swarm participant for many use cases. (Disclosure: Ringer is his product.)
  • Before any self-host conversation, score the 4 conditions honestly — if any fails, price a first-party API or a third-party host in your jurisdiction instead.

Open Questions

  • Whether Kimi K3’s weights actually shipped “very soon” after recording, as Moonshot promised.
  • Which specific seven benchmarks CAISI used, and current DeepSeek pricing vs the May figures — he flags the numbers as already-moving.
  • The evidentiary status of Anthropic’s fraudulent-accounts numbers — described in the video as serious allegations the White House memo did not validate.

Graph View

Backlinks