Model Landscape: Three Months, Four New Models, Two Deprecations
Volcengine Ark Coding Plan has added four new models and deprecated two since June, completely reshaping the lineup. There are now 12 models plus Auto routing—a far cry from the handful available last year. The good news: external model quality has meaningfully improved. GLM-5.2 and V4-Flash GA are a significant upgrade over their predecessors.
Here's the timeline (source):
| Date | Event | Notes |
|---|---|---|
| Aug 7 | DeepSeek V4-Flash GA fully launched | Major agent capability enhancements, no longer preview |
| Aug 4 | MiniMax M2.7, Kimi K2.6 deprecated | Tools using these will stop working |
| Jul 23 | Doubao Seed 2.1 Turbo launched | Balanced performance and cost, upgraded Coding & Agent |
| Jul 17 | Kimi K3 launched | Replaces K2.6, comprehensive coding and reasoning upgrades |
| Jun 17 | GLM-5.2 launched | Zhipu's latest flagship, 1M context |
The 12 models fall into two groups:
Doubao in-house (6): Seed 2.1 Turbo, 2.0 Pro, 2.0 Code, 2.0 Lite, Seed Code, Auto
Third-party (6): GLM-5.2, Kimi K3, Kimi K2.7 Code, MiniMax M3, DeepSeek V4-Flash, V4-Pro
Doubao In-House: Which Model for Which Job?
Auto — Start Here If You Are New
Configure ark-code-latest for Auto routing; the platform picks the best model for quality and speed. You don't need to think about model selection—it handles everything while you code. Perfect for beginners and routine development.
The downside: you won't know which model served which request, so fine-grained cost control is off the table. But honestly, most developers don't need to care.
Seed 2.1 Turbo — The Workhorse
Doubao's best value pick is 2.1 Turbo. 256K context, stronger coding and agent capabilities than 2.0 Pro, yet faster and cheaper. It is the officially recommended workhorse for Coding Plan. Daily completions, code reviews, small refactors—Turbo handles all of it.
If your workflow doesn't involve 50-step agent chains, 2.1 Turbo is the safest default.
Seed 2.0 Pro vs Code: Generalist vs Specialist
2.0 Pro is the general flagship; Code adds multimodal programming chops. Pro excels at complex reasoning and long-chain agents—think requirements doc to PRD to code skeleton in one pass. The Code variant has stronger visual understanding for frontend generation and UI tasks.
For everyday full-stack work, Pro and Turbo cover similar ground. The Code variant shines when prototyping frontends or converting designs to code.
Seed 2.0 Lite / Seed Code — Fillers
Lite is fast and lightweight for autocomplete and low-complexity tasks. Seed Code is the previous-gen code model focused on debugging and refactoring. You will rarely need either unless you want to A/B test Doubao versions on specific tasks.
External Models: GLM-5.2 vs DeepSeek vs Kimi vs MiniMax
GLM-5.2 — The 1M Context Champion
If you are working on codebase-scale tasks, GLM-5.2's 1M context window is the killer feature. That's roughly 150,000 lines of code in one shot (source). Repository-level refactors, cross-module dependency tracing, project-wide code reviews—other models may truncate at their context limit; GLM-5.2 won't.
It also has the best long-horizon performance among external Coding Plan models. The trade-off: higher deduction multiplier. If your Lite plan gives you 18,000 calls per month, running exclusively GLM-5.2 burns through quota faster than Turbo. Save it for heavy lifting.
DeepSeek V4-Flash GA vs V4-Pro
V4-Flash GA is a major leap over the preview version. The official announcement states significant agent capability gains, with benchmarks far exceeding V4-Pro Preview (source). 1024K context, thinking on by default.
The Flash vs Pro decision is simple: Flash for volume, Pro for precision. Unless your task demands extreme accuracy, Flash is good enough and saves on deduction. Reserve Pro for scenarios where mistakes are unacceptable.
Kimi K3 vs Kimi K2.7 Code
K3 is Moonshot's latest flagship; K2.6 is gone. K3 brings comprehensive coding and reasoning upgrades with 256K context, plus more stable long-context and multi-step tool use than K2.6 (source).
K2.7 Code is the coding-optimized variant. If your primary use case is code generation and debugging, K2.7 Code may fit better than K3. If you want an all-rounder, go with K3.
MiniMax M3 — M2.7's Replacement
M3 directly replaces M2.7, ranking at the top of coding and agent benchmarks. 512K context, strong agent reasoning and long-context performance. If you previously used M2.7, switch to M3.
Model Recommendations by Scenario
TL;DR: Daily → Turbo. Heavy lifting → GLM-5.2 or V4-Pro. Frontend → Code. Unsure → Auto.
| Scenario | Recommended Model | Why |
|---|---|---|
| Daily completions/reviews/small refactors | 2.1 Turbo | Fast, cheap, stable—the workhorse |
| Repository-level refactors | GLM-5.2 | 1M context, best long-horizon performance |
| Complex agent orchestration | GLM-5.2 / V4-Pro | Stable long-chain execution |
| Frontend/UI development | 2.0 Code | Multimodal vision + frontend generation |
| Rapid prototyping/high throughput | V4-Flash / 2.1 Turbo | Fast, low deduction cost |
| High-precision critical tasks | V4-Pro / GLM-5.2 | Reliability on hard problems |
| Beginner / hands-off | Auto | Let the platform decide |
Caveats
- Deduction multipliers: GLM-5.2 and V4-Pro have higher deduction rates than Doubao in-house models. The same request count burns through quota faster with these models—Lite plan users should pay attention.
- Auto is not a silver bullet: Auto routing works well for simple tasks but complex agent workflows benefit from manual model selection to avoid mid-workflow switches breaking continuity.
- K2.6 and M2.7 are gone: If you configured these model IDs in Claude Code or Cursor, they will now return errors. Manually switch to their replacements.
- Coding Plan vs API billing are separate systems: Coding Plan bills by request count; API bills by token. Backend services must use API pay-as-you-go, not Coding Plan keys.
- 75% off for first two months: New Lite (
¥9.9) and Pro (¥49.9) subscriptions enjoy roughly 75% off for the first two months, reverting to list price in month three. A great way to deep-test these models before committing to Pro long-term.