Direct Answer
Qwen3.8 Max looks like a capable frontier model with a high ceiling, but this test does not establish it as the best open model. Bijan Bowen's strongest result was genuinely impressive: a detailed C++ skate simulation set on an early-2000s New York block. The wedding website also combined attractive design with coherent long-form writing. But the model missed core CAD constraints, struggled with reference-image fidelity, produced uneven game logic, and needed repair passes to rescue several builds.
The word open also needs a date attached. Qwen's 3 August announcement says the 2.4T Max model and a 27B companion will receive open weights next week. At publication time, the hosted model is available, while downloadable weights, the final model card, file sizes, and license still need to appear. A promise to publish open weights is meaningful. It is not the same as weights readers can download today.
Watch the Full Test
Video credit: Bijan Bowen. Watch the original video on YouTube. Test outcomes and the reported $32 spend are creator observations. Availability was checked against Alibaba Cloud documentation and Qwen's official 3 August announcement.
What Is Actually Released?
| Item | Status on 3 Aug 2026 | Practical meaning |
|---|---|---|
qwen3.8-max | Listed in Alibaba Cloud's current model index | Hosted access is available through regional Model Studio endpoints. |
qwen3.8-max-preview | Still documented on Token Plan and integration pages | Older subscriptions and examples may retain the preview ID. Check the console for the correct regional endpoint. |
| Qwen3.8 Max weights | Announced for the following week | No local deployment claim should be treated as complete until files and a license are published. |
| Qwen3.8 27B weights | Announced for the following week | This is likely the more practical local release, but capability and memory claims remain provisional. |
| Open-source status | Not established | Open weights describe downloadable parameters. Full open source would also require broader training and development artifacts. |
Alibaba Cloud's current model index lists qwen3.8-max for OpenAI-compatible, Anthropic-compatible, and DashScope access in Beijing, Singapore, Tokyo, Frankfurt, and Virginia. Its July Token Plan launch page still describes qwen3.8-max-preview as the latest flagship. That documentation lag is a good reason to copy the model ID from your own console instead of a third-party setup guide.
Specs, Benchmarks, and What Is Still a Claim
Qwen's launch material, as shown in the video and linked from its official announcement, describes a 2.4-trillion-parameter Mixture-of-Experts model with about 95 billion parameters active per token. It also presents a one-million-token context window, up to 128K output, image and video understanding, and benchmark results aimed at frontier coding and knowledge-work models.
Those numbers are useful launch claims, but they are not interchangeable with independent evidence. A parameter count does not measure usability. A benchmark score does not tell you whether a printable motor cavity is correct. A million-token context does not guarantee that an agent will preserve the right constraints across a 50-minute build. The video is valuable precisely because it reveals behavior the headline table cannot.
| Launch claim | How to read it |
|---|---|
| 2.4T total parameters | Model scale and storage burden, not proof of better outcomes. |
| About 95B active per token | MoE compute path; it does not mean only 95B weights must be stored. |
| 1M context | Large input capacity; retrieval quality and long-run instruction retention still need task-specific tests. |
| 128K maximum output | A ceiling, not a sensible default budget for an agent task. |
| 56.6 on DeepSWE 1.1 | A vendor-reported launch result discussed in the video, not a substitute for your repository's acceptance tests. |
| 10+ days of autonomous coding | An endurance demonstration. Reliability depends on the harness, tools, checkpoints, recovery logic, and verifier. |
How Bijan Ran the Test
This is a broad capability probe, not a controlled head-to-head benchmark. Bijan used Qwen's web interface for some one-shot tests and an OpenAI-compatible Qwen endpoint inside Codex CLI for longer builds. He cleared prior cache and memory, generally used fresh sessions, and selected a high reasoning setting. Several projects received follow-up repairs; at least one began in the web app and was then moved into Codex.
Those choices make the video excellent for discovering failure modes and weak for declaring an overall winner. The projects had different time budgets, different repair opportunities, and different harnesses. Some prompts are recurring public tests. A model can also look better in a harness that compiles, opens a browser, inspects screenshots, and retries than it does in a one-shot chat window.
The Ten-Test Scorecard
| Test | What worked | What failed | Verdict |
|---|---|---|---|
| Browser OS | Themed apps, games, sound, personalization, and a memorable window effect | Drawing cursor offset, one-open-only apps, weak GTA-like game | Mixed-positive |
| C++ NYC skate sim | Rich city, cabs, pedestrians, strong atmosphere, repaired ollie and camera | Initial movement, Z-fighting, and camera issues; needed a second pass | Standout |
| Printable RB26 shell | Detailed engine-like form and readable twin-cam treatment | Wrong turbo orientation, no valid motor cavity, ignored support-free print constraint | Failure |
| Street Yeet game | Funny cutscene, generated dialogue, interesting ragdoll effects | Empty first city, invisible entities, incomplete world and weak gameplay | Below target |
| Luxury watch site | Cinematic camera movement and a presentable product page | Band orientation, backward time, static hands, blurry glass | Visually strong, mechanically wrong |
| Wedding writing site | Coherent story, recurring book motif, full event details, attractive palette | No major issue surfaced in the walkthrough | Strong |
| Subway FPS | Competent scene after repair | Buggy web output, generic result, misplaced subway car | Mixed |
| Humor | A serviceable kale-and-chicken joke | One prompt cannot establish durable comedic ability | Acceptable |
| C++ Chrono City | Attempted the requested multi-era concept | Long run ended in a poor, hard-to-use result | Failure |
| Image-to-computer build | Interactive model with spinning fan, LEDs, and auto-orbit | Wrong color, shape, and reference fidelity | Cool generic object, failed reconstruction |
Why the New York Skate Build Matters
The skate simulation is the strongest evidence in the episode because it was not just a polished landing page. The model had to create C++ code, a navigable environment, a player controller, camera behavior, street life, and visual atmosphere. The first pass was broken, but a fresh repair thread recovered the ollie, reduced Z-fighting, improved the camera, and made the scene feel inhabited. The final project exceeded 2,000 lines and was the one result Bijan treated as a clear winner.
That recovery is more informative than a perfect screenshot. It suggests Qwen3.8 Max can inspect a concrete failure, preserve a substantial codebase, and make a meaningful second pass. For production use, though, the success criterion should be automated: the game launches, controls map correctly, collision tests pass, the frame rate stays above a threshold, and a browser or vision check confirms the expected states.
The Failure Patterns Were Consistent
- Spatial plausibility over exact geometry. The watch and computer looked impressive until compared with the requested object. The CAD engine looked detailed while failing the motor-fit requirement.
- Presentation over mechanics. Several games had a convincing mood but weak controls, incomplete worlds, or broken state.
- Constraint loss over long runs. The printable shell ignored core manufacturability instructions even after almost an hour of work.
- Repairability without first-pass reliability. Codex improved multiple outputs, but the cost and time of those repair loops belong in the score.
- Generic substitution. When reference matching became difficult, the model produced an independently attractive object rather than the specified one.
What Did the Full Test Cost?
Bijan reports approximately $32 for the experiments shown. He also discovers that he used a regular Model Studio API key rather than the dedicated Token Plan key. Alibaba's Token Plan documentation says those keys are separate and that the subscription key uses the sk-sp- format. The lesson is operational: confirm which balance an agent is charging before starting parallel 30-to-50-minute builds.
The video cites a pay-as-you-go rate of $2 per million input tokens and $6 per million output tokens. Treat that as the price observed or displayed during the test, because regional billing pages and model status can change. Alibaba's international Token Plan currently advertises limited-time tiers at $6, $18, and $68 per month, with rolling five-hour and seven-day credit limits rather than an unlimited bucket.
The useful metric is cost per accepted result. The $32 endpoint bill excludes the creator's review time, the failed CAD run, the long Chrono City run, and the follow-up repairs. A model that costs less per token can still cost more per usable deliverable if the acceptance rate is low.
Local Hardware Reality: 2.4T Is Not a Laptop Model
Mixture-of-Experts makes inference more efficient by activating only part of the network for each token. It does not remove the need to store the experts. At the announced 2.4T scale, raw weight memory is roughly 4.8 TB at 16-bit, 2.4 TB at 8-bit, or 1.2 TB at 4-bit. Runtime buffers, KV cache, routing, and serving overhead come on top.
| Announced model | 16-bit raw weights | 8-bit raw weights | 4-bit raw weights | Likely audience |
|---|---|---|---|---|
| Qwen3.8 Max, 2.4T | About 4.8 TB | About 2.4 TB | About 1.2 TB | Distributed inference providers and specialist labs |
| Qwen3.8 27B | About 54 GB | About 27 GB | About 13.5 GB | High-memory desktops and consumer local-AI setups after quantization |
These are arithmetic estimates, not deployment recommendations. The 27B model is the one local builders should watch, but its real context capacity, quantization quality, speed, multimodal support, and license cannot be assessed before the release files arrive. Check the official Qwen Hugging Face organization and QwenLM GitHub organization, not a renamed community upload.
A Better Qwen3.8 Evaluation Protocol
- Freeze one real task. Use a private bug, feature, design repair, or document workflow from your own work.
- Keep the harness constant. Same Codex version, tools, permissions, machine, repository commit, and system instructions for every model.
- Write acceptance checks first. Include unit tests, browser assertions, image-difference thresholds, mesh rules, or dimensional tolerances.
- Start clean sessions. Do not let a failed attempt teach the next model where the hidden problem is.
- Equalize repair budgets. Give every model the same number of retries, maximum wall time, and token allowance.
- Track accepted-result economics. Record API cost, elapsed time, human repair minutes, verifier cost, regressions, and final pass rate.
- Repeat the run. One-shot variability is real. Run at least three trials before calling a model reliable.
- Wait for the license. For local or commercial deployment, verify the exact checkpoint, model card, and license after weights are published.
Claim Audit
| Claim | Assessment | Why |
|---|---|---|
| Qwen3.8 Max is the best open model. | Not established | The test is mixed, and the announced weights were not yet downloadable at publication time. |
| It is a 2.4T MoE with about 95B active. | Launch claim | Published in Qwen's announcement material; a full public model card and weights remain the stronger artifact. |
| Open weights arrive next week. | Official dated commitment | Qwen's 3 August post names both Max and 27B. Verify again after the promised window. |
| It can code autonomously for more than 10 days. | Vendor demonstration | Endurance is useful, but the harness, verifier, checkpoints, and intervention rate determine reliability. |
| The model is excellent at C++. | Plausible, task-specific | The skate build was exceptional; Chrono City failed. One category contains both outcomes. |
| It is strong at CAD. | Not supported here | The engine shell missed the defining fit and printability constraints. |
| The full test cost $32. | Creator-reported | A useful observed bill, but not a normalized price-performance benchmark. |
| Max will run locally because it is MoE. | Misleading | Active compute is lower than total size; raw quantized weights remain around terabyte scale. |
Video Chapters
| Time | Chapter | What to watch for |
|---|---|---|
| 00:00 | Intro | Release framing and open-weight promise. |
| 01:18 | First look | Availability, model family, and headline positioning. |
| 03:54 | Technical look | Architecture, context, benchmarks, and API price. |
| 05:20 | Browser OS | Feature breadth versus interaction bugs. |
| 10:16 | C++ skate simulation | The strongest build and its repair pass. |
| 15:01 | 3D CAD | Visual detail versus failed engineering constraints. |
| 18:19 | Street Yeet | Cinematic flair and incomplete gameplay. |
| 22:23 | Luxury watch website | Presentation quality and geometric errors. |
| 24:00 | Creative writing website | The second clear success. |
| 30:19 | Subway FPS | Web failure, Codex repair, generic result. |
| 31:57 | Humor | A tiny and low-confidence writing probe. |
| 32:35 | Chrono City C++ | A long run that failed to become usable. |
| 34:07 | Multimodal coding | Attractive generic output, weak reference fidelity. |
| 35:14 | Results overview | The creator's mixed final ranking. |
| 37:14 | Cost | The $32 bill and wrong-key discovery. |
| 38:24 | Closing thoughts | Why the 27B checkpoint may matter more locally. |
Bottom Line
Qwen3.8 Max deserves serious evaluation, not a coronation. It produced one of the better C++ prototypes in Bijan Bowen's model series and a convincing writing-led website. It also delivered an unusable CAD result, a failed Chrono City build, weak reference reconstruction, and several visually polished projects with mechanical mistakes.
The most important next event is not another reaction video. It is the promised publication of the Max and 27B weights, model cards, and licenses. Once those land, teams can measure reproducible self-hosted performance, serving cost, quantization loss, and commercial terms. Until then, use the hosted model as a candidate generator with strong verification, and judge it by cost per accepted result.
Sources and Further Reading
- Bijan Bowen: Qwen3.8 Max Is HERE - Is THIS the BEST Open Model Yet?
- Qwen: official 3 August Qwen3.8 Max announcement
- Alibaba Cloud Model Studio: current model index
- Alibaba Cloud: Token Plan and Qwen3.8 Max Preview launch
- Alibaba Cloud: Token Plan Personal Edition documentation
- QwenLM official GitHub organization
- Qwen official Hugging Face organization