Qwen had a loud week, and the noise came from three different directions at once. Alibaba pushed its most capable model yet into a limited preview without publishing the paperwork that developers actually need. A lone engineer stuffed a working RISC-V processor into a 4B model and booted Doom inside the weights. And a crypto company with a hardware obsession shipped a 0.8B agent that claims it can out-tool a Qwen model two and a half times its size.
Three stories, one model family, and three very different bets on where local AI goes next. Here is the short version of each, plus what to watch before the month ends.
Qwen4-Max-Preview Is Live, and the Paperwork Is Not
Alibaba opened Qwen4-Max-Preview on November 11 at 10:00 UTC+8, and the official changelog is blunt about its shelf life: the preview runs until 24:00 on November 25, then retires. That gives teams fourteen days to poke at a flagship they cannot download, license, or benchmark against anything published.
The capability list is broad. The model accepts text, images, and video, and the changelog advertises reasoning, web search, tool calling, and structured outputs. Early access also runs through QwenWork, Qoder, the Qwen app, and QwenStudio, so the preview is as much a distribution test as a model test.
What is missing matters more than what shipped. As of November 12, Alibaba had not published parameter counts, weights, a model card, a license, or benchmark scores for the release. Access runs through the Token Plan, and a secondary report lists personal tiers at $6 for Lite, $10 for Essential, $18 for Standard, and $43 for Pro per month, with half-price introductory pricing on the two cheapest tiers. Those are access prices, not per-token rates, which makes cost comparisons against Qwen3.8 guesswork.
The family picture is equally thin. Four names surfaced at the October 29 Apsara Conference: Qwen4-Max, Qwen4-Plus, Qwen4-Flash, and Qwen4-27B. Alibaba Cloud CTO Feifei Li reportedly told the room the family would launch "next week," a window that opened November 2 and had still not produced a full launch when the Max preview appeared nine days later.
- Confirmed: Qwen4-Max-Preview exists, is multimodal, and expires November 25.
- Reported but unconfirmed: a 1-million-token context window, and the four-tier Token Plan pricing.
- Not published: weights, license, parameter counts, benchmarks, and open-weight commitments for any Qwen4 variant.
For anyone building on Qwen today, the practical read is simple. Treat the Max preview as a timed demo, not a platform. If your product depends on self-hosting, the variant to watch is Qwen4-27B, the only name in the lineup that carries its own parameter count. Whether it ships with open weights is still an open question, and that question is the whole ballgame for local deployments.
Qwenputer Puts a RISC-V CPU Inside a 4B Model
While Alibaba kept its flagship behind an API, AI engineer Divyansh Agrawal went the other way and put the computer inside the model. Qwenputer takes stock Qwen 3.5 4B and compiles a 32-bit RISC-V processor directly into the weights. The embedded core runs arithmetic exactly, executes C and MicroPython without an external runtime, and reads and writes through the model's own DeltaNet state, which doubles as RAM.
It also runs Doom. Agrawal demonstrated the E1M1 level on the embedded core, and the frame rate is best described as archaeological. The build even ships a filesystem in the DeltaNet memory: 1,024 slots holding up to 512 files, persistable across sessions by saving model state.
The cost is size. Qwenputer grows to roughly 19GB, nearly double the base model, and Agrawal recommends at least 24GB of VRAM. The code is MIT licensed, with source on GitHub and weights on Hugging Face. It is a stunt, but a pointed one: instead of asking a tool-calling loop to run code outside the model, the execution environment lives in the parameters themselves.
Tether's 0.8B BotPsy Takes a Swing at Qwen
The third story is a straight-up challenge. Tether AI Research released QVAC BotPsy-0.8B on November 11, an open-weight model built for multi-step tool calls on personal devices. In Tether's own tests it scored 44.0 against 37.7 for a Qwen model roughly two and a half times larger, which the company framed as a size-class upset.
The fine print is where it gets interesting. Qwen's published score for that same model narrows the gap considerably, so the headline number depends heavily on whose harness is doing the measuring. Benchmarks run by the vendor releasing the model are marketing until someone independent reproduces them, and tool-use evaluations are especially sensitive to prompt scaffolding and retry budgets.
Still, the direction is hard to ignore. A 0.8B model that handles multi-step tool calls offline is the kind of thing that makes an on-device agent plausible on hardware people already own, no cloud round trip and no per-call bill.
What to watch before November ends. November 25 is the Qwen4-Max-Preview retirement date, and it will tell us whether Alibaba converts a timed demo into a shipped model with a license attached. The Qwen4-27B variant is the open-weight tell, since a locally runnable 27B would do more for the ecosystem than any API tier. And BotPsy needs an independent reproduction before its 44.0 means anything beyond a press release.
For now, Qwen's week reads like a family splitting along three axes at once: closed flagship, embedded computation, and tiny specialists. The interesting part is not which one wins. It is that all three are shipping in the same fourteen-day window.
Comments