跪拜 Guibai
← All articles
AI Coding

Qwen 3.8 Max Ties Kimi K3 in Coding Benchmarks, Then Self-Verifies Its Own Work

By 子昕AI编程 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A model that self-verifies — opening a browser, walking through the UI, and fixing its own bugs — changes the trust calculus for delegating real engineering tasks. Benchmarks don't capture this behavior, but it's the signal a backend engineer actually needs before handing off work.

Summary

A head-to-head rerun of three coding challenges — a video workstation, a real-time seat-selection ticketing system, and a physics driving game — shows Qwen 3.8 Max tying Kimi K3 on features while exceeding it on engineering rigor. The model delivered atomic seat-locking with idempotent payments, a 24-hour waitlist anti-starvation mechanism, and 28 passing unit tests with zero TypeScript errors. In the game scenario, it explicitly adopted planck.js and tuned per-level physics parameters for suspension, low gravity, and ice friction.

The standout behavior was self-verification: after each build, the model opened a browser, walked through the UI, and debugged its own issues — fixing a rAF-throttled preview and a broken Delete-key handler without human intervention. This self-QA step added time but produced deliverables with zero console errors across all three projects.

Qwen 3.8 Max is a 2.4-trillion-parameter MoE with 95B activated parameters and a 1M-token context window, priced at $2/$6 per million input/output tokens. Weights are scheduled for open release, a first for the Qwen Max tier. The test used Qoder in YOLO mode with a consolidated one-shot prompt, giving Qwen a cleaner spec than K3 received, so the tie carries that caveat.

Takeaways
Qwen 3.8 Max tied Kimi K3 on features across a video workstation, a ticketing system, and a physics driving game, and surpassed it on engineering rigor.
The model self-verified each project by opening a browser, testing the UI, and debugging issues like a rAF-throttled preview and a broken Delete-key handler without human help.
All three projects shipped with passing test suites (28, 25, and 23 tests respectively), zero TypeScript errors, and zero browser console errors.
The ticketing system implemented atomic seat-locking, idempotent payment keys, duplicate-verification 409 responses, and a 24-hour waitlist anti-starvation mechanism.
The driving game used planck.js with real per-level physics tuning — suspension springs, collision impulses, low gravity, and ice friction — and included headless simulation playthroughs of all five levels.
Qwen 3.8 Max is a 2.4T-parameter MoE with 95B activated parameters, a 1M-token context window, and API pricing of $2/$6 per million input/output tokens.
Weights are scheduled for open release next week, alongside a 27B small model — the first time a Qwen Max-tier model has opened its weights.
The test gave Qwen a consolidated one-shot prompt derived from K3's multi-round requirements, so the tie should be discounted for the cleaner spec advantage.
New Qoder users get 1,100 free calls (800 on registration plus 300 after login), with off-peak discounts bringing the model to 0.25x Credit.
Conclusions

Self-verification — a model opening a browser, testing its own output, and fixing bugs — is a capability that no benchmark table measures, yet it directly determines whether an engineer can trust the model with unsupervised work.

The pattern of dismissing Chinese models based on outdated impressions has now been broken three times in one year (GLM 5.2, K3, Qwen), suggesting the pace of improvement is faster than community perception updates.

Framework-coupled testing (Qwen+Qoder vs. K3+Kimi Code) is a more honest comparison than bare-model benchmarks because it reflects the actual workflow developers use, not an artificial isolated evaluation.

Qwen's engineering instincts — atomic operations, idempotency keys, 409 conflict responses, anti-starvation queueing — go beyond 'making a demo run' and into production-adjacent thinking about state consistency and failure modes.

The consolidated one-shot prompt gave Qwen an unfair advantage over K3's multi-round process, but the self-QA behavior and test coverage suggest the gap in raw capability may be smaller than the prompt advantage implies.

Concepts & terms
MoE (Mixture of Experts)
A model architecture where only a subset of parameters (experts) are activated per token, reducing compute cost. Qwen 3.8 Max has 2.4T total parameters but activates only 95B per inference.
SSE (Server-Sent Events)
A unidirectional server-to-client push protocol over HTTP, used here to stream real-time seat status updates to the browser without polling.
Idempotency key
A unique identifier sent with a request that ensures duplicate submissions produce the same result — critical for payment endpoints to prevent double-charging on retries.
planck.js
A JavaScript physics engine ported from Box2D, used for 2D rigid-body simulation with collision detection, joints, and forces — the engine Qwen explicitly adopted for the driving game.
rAF throttling
Browsers throttle requestAnimationFrame callbacks in background tabs to save resources, which can freeze animations or previews. Qwen detected this and added a setInterval fallback.
YOLO mode
An autonomous agent mode in Qoder where the model writes code, runs commands, and iterates without asking for human approval at each step.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗