Qwen4 Leak: 20-Line PR Sets 1M Context, Reveals 120B Multimodal Model

Three Timelines Pointing to Qwen4

As of October 28, three signals converge on Qwen4: the Amsterdam Apsara conference where Qwen occupied two agenda slots, a countdown page scheduling release for October 30, 2024 02:00 Beijing time (2026-10-29 10:00 UTC-8), and a merged pull request in the QwenLM/qwen-code repository that configures the entire Qwen4 generation for a 1 million token context window. Official channels have not yet announced Qwen4.

qwen.ai Research Page: No Qwen4

The qwen.ai research page's "Latest Progress" section lists four entries dated September 20 and 18, 2024 — Qwen-Image-2.1, Qwen3.8-LiveTranslate, Qwen3.8-Omni-Flash, and E-Commerce Bench — all belonging to the Qwen3.8 series. No Qwen4 entries appear.

The 20-Line PR: What It Changed

PR #15762, authored by wenshao and merged on October 28, carries the title fix(core): size the qwen4 generation at 1M. It modifies two files with a net +20 lines. The change adds a single rule for the Qwen4 family, placed beside the existing Qwen3.x rule and before the 256K fallback line. This rule matches all naming variants — bare major version, minor versions, coder variants, dated snapshots, and provider-prefixed IDs — and assigns them a 1,000,000 token context window, eliminating the need to update code for each new release.

Before this fix, the models.dev registry and internal directory contained zero Qwen4 entries, causing the entire generation to fall back to the 262,144 (256K) default. The symptom was subtle: the client believed the window was only one-quarter of the true size, triggering automatic compression at 22% of the real window without any error.

The PR includes a verification list showing eight IDs that previously all resolved to 262,144: qwen4, qwen4-max, qwen4-plus, qwen4.5-max, qwen4-coder-plus, qwen4-max-2026-10-01, openai/qwen4-max, and qwen4-omni-120b-think. The author notes the last ID is already used as a session model in Omni test fixtures, confirming these IDs are not hypothetical. A 120B multimodal Qwen4 has been running internally for more than a day.

Verification method: connect an OpenAI-compatible provider, request /context with any Qwen4 ID; before the fix it returns 262.1k, after it returns 1000.0k. The same "pin new generation once" convention was previously applied to Gemini 3.x, GLM-5.2+, and Kimi K3.

Arena's Kerria: 32 Minutes vs GPT-6 Astra's 49 Minutes

On October 27, an anonymous model codenamed Kerria appeared on LMArena. Tokenizer traps identified it as Qwen4: four tokens that only Qwen-series models stumble on, Kerria stumbled on three; 52 tokens from 13 other providers (Kimi, DeepSeek, OpenAI, Google, Mistral, etc.) caused zero stumbles. Previous K-codenamed models — Kiteki, Kizen-alpha, Korin — corresponded to Qwen3.5, 3.6, and 3.7 Max previews.

A head-to-head test on the same prompt — an interactive Three.js neon Kowloon Walled City with five layers in a typhoon night — showed Kerria (Qwen4) completing in 32 minutes versus GPT-6 Astra's 49 minutes. Observers noted Astra's structure was cleaner with more interactive options, while Qwen4's output was denser. The 17-minute gap does not prove overall superiority but demonstrates this generation can handle long-horizon tasks.

The Countdown Page: Unofficial

A countdown page screenshot circulated on October 28 18:00, displaying "Qwen4 Open Weight Soon" and a release time of 2026-10-29 10:00 UTC-8 (October 30, 2024 02:00 Beijing time). The page mimics official style but @Alibaba_Qwen has not retweeted it; the poster even asked "did you forget to post this?".

Verdict: Code Proves Internal Use, Not Imminent Release

Among the three signals, the code is the hardest evidence — but it proves the model is already in use internally, not that a public release is imminent. A fully multimodal model that exists only internally and is absent from the registry was written into test fixtures; a PR pins the entire generation at 1M context, covering even qwen4.5-max. Such changes are not made for a non-existent model.

LMArena's Kerria shows the model can produce working output, but anonymous arena battles are not a release. The countdown page is the softest signal: no official endorsement, and the time is set to US Pacific morning.

Official statements so far consist of two points from the September 22 Yunqi conference: Qwen4 adopts a next-generation architecture and is in training; subsequent Qwen4.5 and Qwen5 are planned to scale to 5–10 trillion parameters. The four tiers — Max, Plus, Flash, 27B — appeared only on a main-forum slide, not in written press releases. The 27B tier is positioned as the open-weight release, but its license, parameter count, pricing, and benchmarks remain unpublished.

The only definitive switch will be the appearance of a Qwen4 repository on Hugging Face or a qwen4 model ID in API documentation. Until then, the latest downloadable release remains Qwen3.8-27B.

Implications for Local Deployment

For local deployment: 256GB RAM is sufficient to run the 27B open-weight model with headroom (Qwen3.8-27B Q4 quantization ≈ 16GB). The 120B multimodal tier will likely require multi-machine or multi-GPU setups; exact weight sizes must wait for release. The 20-line PR tells us Qwen4 is on the way, but the countdown page should not be treated as an alarm clock.

Data sources: QwenLM/qwen-code PR #15762 (public repository), LMArena anonymous model Kerria tracking, community-shared Apsara countdown page screenshots; official statements from the September 22 Yunqi conference. Model IDs, line counts, and timestamps are taken directly from the disclosed information without extrapolation.