Ripple 65B Preview vs Qwen 4 Max: One of These Models You Can Actually Call

A comparison page needs two models. For this one, only one exists. Ripple 65B Preview shipped on 4 November 2026 as a 65-billion-parameter dense model under Apache 2.0, with weights on Hugging Face and a third-party API listing. Qwen 4 Max has been a name on a slide since Alibaba's Apsara conference on 22 September 2026, and as of 5 November there is no model card, no weights, no price, no API identifier and no published benchmark for it anywhere. Every number you have seen attached to Qwen 4 Max this week is a projection, including some that were presented as fact.

That does not make the comparison useless. It makes it a different comparison: the shipped model against the announced one, and the more practical question of what you should run today while the announced one stays announced. For that, the honest stand-in is Qwen3.8-Max — Alibaba's current closed flagship, live since 3 August, and the model Qwen 4 Max will actually replace.

What Alibaba has actually said, in order

On 22 September 2026 at Apsara in Hangzhou, Alibaba showed four tier names: Qwen4-Max, Qwen4-Plus, Qwen4-Flash and Qwen4-27B. The official press release said the family "is currently in training." It gave no date, no parameter counts, no context window, no licence and no price. The tier names themselves came from attendees' photographs of the stage, not from the written release — which is worth remembering the next time you see them quoted as a specification.

On 29 October, at Apsara Amsterdam, Alibaba Cloud CTO Feifei Li was reported by attendees and press as saying the family would arrive "next week." Alibaba's own written material still said "in training" at the time. The week of 2 November came and went. On 4 November, Alibaba restated its commitment to open weights for Qwen 4 and Qwen 5 — a statement about intention, not a release.

So the timeline as it stands: a keynote, four names, one executive's timeline that slipped, and two commits to a command-line client. That is the entire evidence base.

The commits are not a spec sheet

The most concrete artefact is Qwen Code pull request #15762, merged on 28 October and shipped in v0.26.3, which sizes every qwen4 identifier at a 1M-token context in the client's defaults. A follow-up, #15987, merged on 31 October and shipped in v0.26.5, gives those identifiers a low/medium/xhigh reasoning-effort ladder with xhigh as the default.

It is tempting to read 1M context out of that as an official figure. It is not. It is a client-side default written by engineers who expected an endpoint to exist, and competing readings of the same table put the value at 262,144 instead. The identifier table also contains no qwen4-27b entry and the Qwen Hugging Face organisation has no 27B repository, even though the 27B tier was named on the Apsara slide. When the tooling is that incomplete, the numbers in it are guesses with a commit hash attached.

This matters because a small industry has grown up around filling the gap. Several aggregator pages now carry full Qwen 4 Max spec tables — parameter counts, SWE-bench scores, per-token prices, endpoint names — with no provenance. Some of them are on domains that look authoritative. None of them trace back to Alibaba. If you are building a procurement case on a Qwen 4 Max number this month, you are building it on something somebody invented.

The substitution that actually answers your question

If what you need is a Qwen flagship to compare against Ripple 65B Preview, the one that exists is Qwen3.8-Max. It is a Mixture-of-Experts model with roughly 2.4 trillion total parameters and 95 billion active, a 1M-token context, a 131,072-token output ceiling, and text, image and video input. It costs $2.00 per million input tokens and $6.00 per million output, and it scores 45.4 on Artificial Analysis's Intelligence Index.

• Parameters — Ripple 65B dense vs Qwen3.8-Max ~2.4T total / 95B active

• Weights — Apache 2.0 on Hugging Face vs closed, API only

• Price — $0.20 / $0.60 per million on the first hosted endpoint vs $2.00 / $6.00

• Context — 1M advertised, 16.4K on one third-party listing vs 1M served

• Input modality — text only vs text, image and video

• Agentic coding — 80.5 on Terminal-Bench 2.1, vendor-reported vs no Terminal-Bench figure published

The pattern repeats what the other comparisons in this series find. Ripple is ten times cheaper on input and ten times cheaper on output, it is the only one of the pair you can download and fine-tune, and it is the only one making a loud agentic-coding claim. Qwen3.8-Max is the one you call when the job needs video input or a context window measured in millions of tokens with a serving guarantee behind it.

What Qwen 4 Max would have to do to be worth waiting for

Four things would make the wait rational, and none of them has happened yet.

First, publish the card. A parameter count and an architecture description are the minimum for anyone to plan capacity, and "four tier names on a slide" is below that bar.

Second, publish a price. The Max tier has been closed-weight for over a year and priced at the top of Alibaba's range. If Qwen 4 Max arrives at the same $2/$6 as its predecessor, the price argument against Ripple does not change; if it arrives above, it gets worse.

Third, settle the licence. Alibaba has restated an open-weight commitment for the Qwen 4 family and named the 27B tier first, which implies the Max tier stays closed. A rumour about revenue-sharing for large commercial users has circulated without confirmation. Until a licence file exists, plan for closed.

Fourth, publish a benchmark at a version someone else has run. Ripple's 80.5 on Terminal-Bench 2.1 is a vendor claim on an older revision, and it is still more than Qwen 4 Max has offered in six weeks of being announced.

Why the blank column is not a reason to wait

There is a real cost to holding a decision open for a model with no release date. If you need an agentic coder this week, the choice in front of you is Ripple 65B Preview at roughly $0.20 and $0.60 per million tokens under Apache 2.0, or Qwen3.8-Max at $2.00 and $6.00 behind a closed API. That is a decision you can make and reverse. Waiting for Qwen 4 Max is not a decision; it is a deferral with an unknown end date, and the CTO's "next week" has already missed once.

The rational posture is to build against what exists and keep the migration cheap. Qwen3.8-Max is on our catalogue today at $2.00 and $6.00 with 0% markup — provider list price passed through, so if Alibaba reprices the tier when Qwen 4 Max lands, the change reaches your key the same day rather than at the next contract. Ripple 65B Preview is not on our catalogue; it reaches an API through Ricursive's own endpoint and third-party platforms, and the weights are on Hugging Face if you would rather run it yourself.

Building against the OpenAI-compatible shape both of them use is what keeps the swap to Qwen 4 Max a configuration change rather than a rewrite — and it is the only hedge available against a release date nobody has published.

What to do on 5 November

If your work is text-only agentic coding and cost is the constraint, run Ripple 65B Preview. It is available now, it is Apache 2.0, and its Terminal-Bench claim is the strongest either lab has made this month — treat 80.5 as Ricursive's own number on an old revision, and reproduce it on your tasks before you commit.

If your work needs image or video input, a guaranteed 1M-token context, or a vendor with a support contract, run Qwen3.8-Max today and put Qwen 4 Max on a watch list. Watch for a model card, not a keynote quote; for a licence file, not a commitment; for an API identifier that resolves, not a client-side default in a pull request.

And when Qwen 4 Max does ship, this page becomes a different article. Right now the honest verdict is narrower than a benchmark table would suggest and simpler than a roadmap: one of these models you can call, and one of them you cannot.