Sunday, 20 September 2026

Alibaba's Qwen team released Qwen-Image-2.1 on 20 September, an open-weight image generation and editing model whose visual generator carries 7 billion parameters and runs on a single consumer GPU such as an RTX 3090, according to [1]. The model natively produces and edits transparent (RGBA) images, generates at 2K resolution, and can take up to 10 reference images at once for tasks like group portraits, virtual try-ons or room layouts, according to [2]. Local edits are guided by circles, masks or painted marks rather than text alone [2].

The team says the 7B generator beats most closed models on its own benchmark, though independent results are not yet public [1]. A score a vendor reports on its own test is not the same as a third-party result, and Qwen concedes as much.

The catch is the licence. Qwen-Image-2.1 ships under the Qwen Research License, which permits research and non-commercial use only; commercial use needs a separate agreement with Alibaba [1][2]. So the weights are free to download and study, but a small business cannot fold the model into a paid product without asking first. Diffusers, ComfyUI, vLLM and SGLang support it from day one [2], so the setup cost is low for anyone already running local image tools.

Chinese lab StepFun announced Step 5 Preview on 20 September and opened API access the same day, according to [1]. The model is a sparse mixture-of-experts design with about 600 billion total parameters and roughly 27 billion active per token, and it ships with a 1-million-token context window aimed at long-running agent jobs [1]. StepFun prices it at $1 per million input tokens and $2.70 per million output tokens, with a cache discount, according to [2].

That pricing is the story for anyone buying model calls. A 600-billion-parameter model at $1 input undercuts most Western frontier APIs, where input on a top model costs several times as much. StepFun pitches Step 5 at software engineering, financial analysis and other multi-step professional work, according to [1] — tasks where a long context and cheap tokens matter more than a leaderboard win.

On Artificial Analysis's independent Intelligence Index, Step 5 Preview scores 44, above average for reasoning models in its price tier, and runs at about 99.8 tokens per second in that test [2]. It accepts both text and images [2]. The word "preview" matters: this is an early release, not a settled product, and behaviour can change before any stable version. For a small team the read is simpler — a cheap, long-context model worth testing on a real workflow before committing.

ChangXin Memory Technologies (CXMT) said on 20 September, at the World Manufacturing Convention in Hefei, that its fifth-generation "G5" DRAM platform has entered mass production, according to [1]. The first products are two 24-gigabit LPDDR5X mobile memory chips that hold 50% more data than the 16-gigabit parts they replace, according to [2].

The technical claim is that CXMT reached an 11.95-nanometer active-array half-pitch using quadruple patterning, and gets at least 50% more usable dies from each wafer than its previous generation [1][2]. A company vice-president, Luo Xiaodong, says the process is "on par with the most advanced nodes now in mass production" [2] — a vendor claim, and one independent teardown analysts have not yet confirmed.

The context is supply. CXMT has scaled from roughly 40,000 wafers a month in 2020 to about 300,000 today, and targets 350,000 to 375,000 by the end of 2026, according to [1]. It completed a Shanghai STAR Market listing in July 2026 that raised about $8.6 billion [1], and now ranks as the world's fourth-largest DRAM maker, with a 9.5% revenue share in the second quarter of 2026 [2]. More Chinese memory output matters because a memory shortage has pushed AI hardware and RAM prices up this year; a credible fourth supplier eases that pressure over time, not overnight.

Vals AI raised a $40 million Series A led by Andreessen Horowitz, TechCrunch reported on 19 September [1]. The company had earlier raised a seed round from 8VC and Bloomberg Beta [1]. Vals builds benchmarks that test models on real professional work — law, finance, coding, cybersecurity and biosecurity — and keeps its test materials private so model makers cannot train against them [1][2].

The pitch is trust. Public leaderboards can be gamed once a test set leaks into training data; Vals sells evaluation the way the College Board sells the SAT, charging model makers to be graded on tasks they cannot see in advance [1]. Founder Rayan Krishnan, 25, says the goal is to measure "the real impacts of the models" and whether they produce work "of the same quality as a human" in each domain [1].

The business is growing on the back of that pitch. Revenue is now eight times last year's level, headcount has gone from 8 to 25 people, and the company plans to hire another 10 to 15, according to [1][2]. It has also started evaluating models for US federal agencies [1]. For a buyer, the useful signal is not the round size but the direction: independent, confidential evaluation is becoming a paid layer between the labs and the companies deciding which model to trust.

President Trump used Truth Social on 19 September to propose renaming artificial intelligence and to say he is forming an "AI Force," TechCrunch reported [1]. Trump wrote that "artificial intelligence" is an inaccurate and ineloquent label and floated "Superior Intelligence," "Extreme Intelligence" or "Supreme Intelligence" as replacements, posting a poll for followers to vote on, according to [2].

On the AI Force, Trump compared it to the Space Force he created in his first term, which he called a success, and said he would name an "AI Czar" in the near future, adding that "only high-IQ individuals need apply," according to [2]. He did not describe what the AI Force or the czar would actually do [1][2]. No budget, staffing or legal authority was announced [1].

For a small operator, none of this changes a tool, a price or a rule today. It is worth noting only because it signals how the US administration intends to talk about and organise around AI — branding and a new office, ahead of any concrete policy. Trump also framed AI as central to US competitiveness against China, according to [2]. Until a mandate, budget or regulation follows, the practical effect on anyone running a business on these tools is zero. The naming poll is a poll, not a policy.