Step 5 vs Kimi K3, GLM-5.3, Qwen and DeepSeek Compared

👁 117 reads

The new Step 5 Preview from StepFun enters an elite tier of Chinese models alongside Kimi K3, GLM-5.3, Qwen3.8, and DeepSeek V4.1. On the independent Artificial Analysis Intelligence Index, it ties Kimi K3 at 44 points and sits just one point behind GLM-5.3, while undercutting both on listed pricing. In our multilingual tests conducted in Polish, however, it shared the bottom spot with GLM. Below is an objective, data-driven comparison covering benchmark indices, actual API costs, context limits, open-weight licensing, and hands-on test runs.

Model Comparison Table

Source for benchmarks, pricing, context windows, and parameters: Artificial Analysis model pages, recorded October 9, 2026. Pricing reflects direct developer API rates; OpenRouter rates vary by routing provider.

Intelligence Benchmarks: A Dead Heat at the Top

The single-point separation across GLM-5.3, Qwen3.8 Max, Step 5, and Kimi K3 effectively amounts to a statistical tie. What stands out is architectural efficiency: Step 5 hits this performance tier activating only 27B parameters, whereas Kimi K3 engages 104B active parameters. That architecture shows up directly on the invoice: StepFun claims a 65% reduction in total task cost versus GLM-5.3 and Kimi K3 run in their maximum reasoning modes.

According to figures reported in industry media from StepFun internal evaluations, Step 5 holds an edge in code generation: 67.7 on DeepSWE versus 67.5 for Kimi K3 and 66.9 for GLM-5.3, alongside 80.5 on ProgramBench compared to 77.8 and 72.0 respectively. On Terminal-Bench v4, it trails GLM-5.3 significantly (33.3 vs 41.9), though it comfortably beats Kimi K3 (12.6). Alibaba and DeepSeek were not included in those specific comparison runs.

API Pricing: Cost Per Token vs Actual Task Expense

Nominal token rates can be deceptive. While Step 5 features an attractive output rate on paper, its reasoning mechanism generates substantial token volume. In our 9-prompt writing evaluation, Step 5 cost six times more than GLM-5.3. Across the complete test batch (factoring in required retries due to empty responses), it ended up ten times as expensive: $0.31 versus $0.03. Artificial Analysis observed the same pattern: in their suite, Step 5 output 160M tokens against an evaluation median of 94M tokens. For agentic pipelines with heavy context, always calculate total run cost against real production inputs rather than raw rate tables.

Multilingual Evaluation: Non-English Benchmarks

We tested each model across 13 tasks in Polish (9 composition prompts, 2 editing exercises, 2 knowledge quizzes), enforcing low reasoning effort and matching maximum token limits. Step 5, Qwen3.8 Max, and GLM-5.3 were judged simultaneously on October 9, while Kimi K3 and DeepSeek were scored in an earlier testing batch using identical anchor-model calibrations. Differences within 2-3 points should be considered within the margin of error.

Step 5 lost significant ground due to three completely blank returns: across an e-commerce product description and both editing tasks, its entire token ceiling was consumed purely by reasoning tokens before writing any final answer. Where it did complete the task, it performed cleanly on concise formats (email drafting rated 8.5/10), but dropped to 5.0 on extended creative fiction. Full test details: Step 5 multilingual testing breakdown.

Weights, Licensing, and Self-Hosting

• DeepSeek V4.1 Flash and V4 Pro: Released under the permissive MIT license, offering the fewest operational and commercial restrictions.
• Qwen3.8 27B: Open under Apache 2.0. At 27B parameters, it fits within 4-bit quantization on a single 24 GB GPU. Qwen3.8 Max remains API-only.
• Kimi K3 and GLM-5.3: Open weights, but governed by custom bespoke licenses carrying specific commercial restrictions and usage disclosures.
• Step 5: Model weights are scheduled for an October 15 release, with license terms still pending. At 600B total parameters, deploying unquantized BF16 weights requires roughly 1.2 TB of VRAM, limiting local hosting to enterprise server racks.

Which Model Should You Deploy?

• For non-English writing, localized copy, and proofreading: Kimi K3 or Qwen3.8 Max offer the strongest multilingual fluency. If budget is the primary constraint, DeepSeek V4.1 Flash is the most cost-effective alternative.
• For API-based coding workloads on a budget: Step 5 Preview or GLM-5.3. Step 5 shows better synthetic benchmark results on DeepSWE and ProgramBench, while GLM holds the advantage in shell and terminal agent environments.
• For long-context multimodal inputs: Step 5, which handles native video files up to 128 MB and pairs a 1M token context window with prompt caching at $0.05 per million tokens.
• For self-hosting without licensing headaches: DeepSeek (MIT) or smaller Qwen models (Apache 2.0). For Step 5, wait for the full license terms on October 15 before planning local infrastructure.

These findings reflect Artificial Analysis metrics alongside our direct testing run. You can compare all of these LLMs directly against other leading frontiers in our Model Arena.

// read moreStepFun Step 5 Preview Hits OpenRouter: Kimi K3 Rival at $1NewsDeepSeek, GLM, Kimi and Qwen: Free AI or Cheap API?AI ToolsLesser-Known AI Models: MiMo, Qwen3.8 Max, GLM-5.3, Muse SparkBenchmarks

Kinetyra