Cloudflare and Microsoft both shipped Qwen derivatives

Two model releases landed on the same day, from two companies that compete with each other, and neither trained the model it is now selling.

Cloudflare released Clef-omni, an open-weight decision model built on Qwen3-Omni-30B-A3B-Instruct that takes audio and video alongside text and images at $0.15 per million input tokens, on its own blog. Microsoft released Microsoft-Decision-1, a closed-weight decision-scoring model post-trained from Alibaba's Qwen3.5-9B, priced at $0.042 per million input tokens with 85 ms median latency, reported by MarkTechPost. One is open, one is closed, one costs nearly four times the other, and both start from weights Alibaba paid to train.

The product is the price and the latency

What each company added is the part it is charging for: hosting, latency, a price. Microsoft reports 83.5% average accuracy across the 36 benchmarks it ran itself, which is a vendor's own test suite rather than an outside result. Cloudflare, releasing the same day, cut Clef-flash to $0.038 and dropped its hosted context window from 64k to 24k in the same move. When price is the axis of competition, capacity is what gives.

That is a market structure, not a one-off. The capability arrives from outside, cheaply enough that two platform companies can reprice it within one day of each other, and the differentiation moves to the serving layer. Which has its own exposure: Nvidia's DCGM Exporter carries an unauthenticated resource exhaustion flaw, CVE-2026-47483 at CVSS 8.2, and Lava Security found more than 2,000 GPU servers exposing the tool to the public internet with no authentication at all, per CSO Online.

A derivative inherits its base

The same openness that makes a base cheap to build on makes it cheap to strip. Tech Against Terrorism rated more than 130 AI models on counter-terrorism safety and found three in five failed, with Meta's Llama 3.1 8B falling from 97 to around 3 out of 100 once abliteration removed its guardrails, reported by CBS News.

That figure is about Llama, not about either model released today. But Clef-omni ships open-weight, which is the property the abliteration result is about: whoever holds the file holds the guardrails. A safety claim made by the company that served you the model is not a claim about the weights once they are on someone else's machine.

Who is still paying to train

Google is the contrast on today's record. It is rolling out Gemini 4 Argon to Fairwind Program partners ahead of a public release while employees test a later model code-named Carbon that they say feels comparable to Claude Opus 5.5 on coding, according to Business Insider. That is in-house training at the frontier, kept inside until it ships. Two tiers are visible in one day's items: a frontier that trains, and a serving tier that post-trains what the frontier and the Chinese labs put out.

One more item on today's record reads differently against that split. A Working Group on AI Pause Feasibility of 26 researchers from UC Berkeley, Princeton, Stanford, Harvard and Oxford concluded that a hardwired global pause on frontier training of at least 10 years is feasible, by halting production of training chips and replacing the existing stock with inference-only hardware, per Berkeley News. The report argues about frontier training. It says nothing about the layer below it. But the market that layer describes, where capability is imported and the work is pricing it, is close to what remains when nobody can train anything new: the bases freeze, and the competition is entirely over who serves them for $0.038.