> This is the markdown version of https://www.maniac.ai/blog/chinese-frontier-models-compared-glm5-minimax-kimi-qwen. Visit the full page for interactive content. # Chinese frontier models compared: GLM-5, MiniMax M2.5, Kimi K2.5, and Qwen 3.5 | Maniac | Maniac [Blog](/blog) February 28, 2026 February 2026 was a landmark month for Chinese AI labs. Within weeks of each other, four models shipped that compete with — and in some cases exceed — Western frontier systems on standardized benchmarks, at a fraction of the API cost. For anyone running AI in production, the model landscape just got a lot more interesting. This article compares **GLM-5** (Zhipu AI), **MiniMax M2.5** (MiniMax), **Kimi K2.5** (Moonshot AI), and **Qwen 3.5** (Alibaba) head-to-head against **Claude Opus 4.6** across coding, math, reasoning, agentic tasks, and multimodal benchmarks. All scores are sourced from [Vals AI](https://www.vals.ai), which runs every model through the same standardized evaluation harness — no cherry-picked lab numbers. ## Quick links - **Benchmark tables** — Jump to Head-to-head benchmarks - **Pricing** — Jump to Pricing comparison - **Decision guide** — Jump to When to use which model - **Inference infrastructure** — See [Inference stacks compared](/blog/inference-stacks-vllm-tgi-tensorrt) * * * ## The models at a glance Model Lab Total params Active params Architecture Context License Release Claude Opus 4.6 Anthropic Undisclosed Undisclosed Dense (assumed) 200K Proprietary Feb 5, 2026 GLM-5 Zhipu AI **744B** 40B MoE 137K MIT Feb 11, 2026 MiniMax M2.5 MiniMax 230B **10B** MoE (256E / 8A) 197K Open weights Feb 12, 2026 Kimi K2.5 Moonshot AI **1T** 32B MoE (384E / 8A) 262K Modified MIT Jan 26, 2026 Qwen 3.5 Alibaba 397B 17B MoE (hybrid attn) **991K** Apache 2.0 Feb 16, 2026 Every Chinese model uses a **Mixture-of-Experts (MoE)** architecture, activating only a fraction of total parameters per token. Kimi K2.5 packs 1 trillion total parameters but activates just 32B per forward pass. MiniMax M2.5 takes efficiency further: 10B active parameters producing competitive SWE-bench scores. This is the mechanism behind the dramatic cost gap. * * * ## A note on methodology All benchmark scores below come from [Vals AI](https://www.vals.ai), which uses a **standardized evaluation harness** for every model. For SWE-bench, this means all models run through the same [SWE-Agent](https://github.com/SWE-agent/SWE-agent) framework — no custom harnesses or optimized prompts. This matters. Labs frequently report higher scores using their own optimized setups. For example, Anthropic reports 80.8% on SWE-bench using a custom harness, while Vals AI measures 79.20% using SWE-Agent. The vals.ai numbers are the apples-to-apples comparison. * * * ## Head-to-head benchmarks Bold scores indicate the best result in each row. Dashes indicate the model was not evaluated on that benchmark by Vals AI. ### Coding Benchmark Claude Opus 4.6 GLM-5 MiniMax M2.5 Kimi K2.5 Qwen 3.5 SWE-bench **79.20%** 67.80% 70.40% 68.60% 70.40% LiveCodeBench v6 84.68% 81.87% 79.21% 83.87% **85.33%** Terminal-Bench 2.0 **58.43%** 49.44% 41.57% 40.45% 41.57% IOI — 22.00% 6.67% 17.67% — _SWE-bench: resolving real GitHub issues. LiveCodeBench: competitive programming. Terminal-Bench: command-line task execution. IOI: International Olympiad in Informatics._ Claude Opus 4.6 leads SWE-bench decisively at 79.20% — 9+ points ahead of every Chinese model. On the standardized harness, the gap is wider than lab-reported numbers suggest. LiveCodeBench tells a different story. Qwen 3.5 (85.33%) actually edges out Claude Opus 4.6 (84.68%), with Kimi K2.5 (83.87%) and GLM-5 (81.87%) also competitive. Competitive programming is a clear strength for Chinese models. Terminal-Bench 2.0 — real-world command-line tasks — remains a stronghold for Claude Opus 4.6 (58.43%). The Chinese models cluster between 40–49%, with GLM-5 performing best among them at 49.44%. IOI (olympiad-level algorithmic problems) shows GLM-5 at 22.00% and MiniMax M2.5 at just 6.67%. Deep algorithmic reasoning remains a challenge for MoE architectures. ### Math & reasoning Benchmark Claude Opus 4.6 GLM-5 MiniMax M2.5 Kimi K2.5 Qwen 3.5 AIME **95.63%** 91.67% 88.75% **95.63%** 86.04% GPQA Diamond **89.65%** 83.33% 82.07% 84.09% 87.37% MMLU-Pro **89.11%** 86.03% 80.09% — 87.18% _AIME: competition math (8 runs averaged). GPQA: PhD-level science. MMLU-Pro: graduate-level knowledge across 14 disciplines._ Claude Opus 4.6 leads on GPQA (89.65%) and MMLU-Pro (89.11%). On AIME, Kimi K2.5 ties Opus at 95.63% — a remarkable result for an open-weight model at a fraction of the cost. Qwen 3.5 is the strongest Chinese contender on knowledge benchmarks, scoring 87.37% on GPQA and 87.18% on MMLU-Pro, within 2 points of Opus on both. GLM-5 (91.67% AIME) and MiniMax M2.5 (88.75% AIME) are solid but not at the same tier. ### Multimodal Benchmark Claude Opus 4.6 GLM-5 MiniMax M2.5 Kimi K2.5 Qwen 3.5 MMMU 83.87% — — **84.34%** 22.77%\* _MMMU: university-level multimodal reasoning._ \*_Qwen 3.5's 22.77% MMMU score (last place, 58/58) likely reflects an evaluation issue rather than actual capability — Alibaba reports substantially higher scores using their own harness._ Kimi K2.5 edges out Claude Opus 4.6 on MMMU (84.34% vs 83.87%) — one of the few benchmarks where a Chinese model takes the lead. GLM-5 and MiniMax M2.5 were not evaluated on MMMU by Vals AI. ### Vals Index (overall composite) The Vals Index is a composite score across multiple enterprise-relevant benchmarks including finance, legal, medical, tax, and coding tasks. Model Vals Index Cost/Test Latency Claude Opus 4.6 **65.98%** $1.00 337s GLM-5 60.69% — — Kimi K2.5 59.74% $0.13 378s Qwen 3.5 57.06% $0.31 575s MiniMax M2.5 53.57% **$0.16** **264s** Claude Opus 4.6 leads the composite at 65.98%. Among Chinese models, GLM-5 (60.69%) and Kimi K2.5 (59.74%) are closest to the frontier. MiniMax M2.5 trails at 53.57% overall but is the cheapest ($0.16/test) and fastest (264s latency). ### Enterprise benchmarks (selected) Vals AI also tests domain-specific enterprise tasks. Here are notable results: Benchmark Claude Opus 4.6 GLM-5 MiniMax M2.5 Kimi K2.5 Qwen 3.5 CorpFin 67.02% 62.90% 59.60% **68.26%** 65.31% TaxEval v2 **75.96%** 70.03% 68.15% 74.20% — Finance Agent **60.05%** 53.18% 38.58% 50.62% 54.47% MedQA **95.41%** 94.27% 92.53% 94.37% 95.21% LegalBench **85.30%** 84.06% 79.96% — 85.10% _CorpFin: corporate finance reasoning. TaxEval: tax law. Finance Agent: multi-step financial workflows. MedQA: medical questions. LegalBench: legal reasoning._ Kimi K2.5 takes #1 on CorpFin (68.26%), beating Claude Opus 4.6. Qwen 3.5 nearly matches Opus on LegalBench (85.10% vs 85.30%) and MedQA (95.21% vs 95.41%). Claude leads most other enterprise benchmarks, but the margins are often slim. * * * ## Pricing comparison Model Input (per 1M tokens) Output (per 1M tokens) Cost vs Opus Claude Opus 4.6 $5.00 $25.00 **1x** GLM-5 $1.00 $3.20 5x cheaper Qwen 3.5 $0.60 $3.60 8.3x cheaper Kimi K2.5 $0.60 $2.00 8.3x cheaper MiniMax M2.5 **$0.30** **$1.10** **16.7x cheaper** MiniMax M2.5 is **16.7x cheaper** than Claude Opus 4.6 on input tokens and **22.7x cheaper** on output tokens. Even GLM-5 (the most expensive Chinese model) is 5x cheaper than Opus on input. ### What this means at scale Scenario: 1M calls/day (1K tok/call) Claude Opus 4.6 GLM-5 Kimi K2.5 MiniMax M2.5 Daily cost $5,000 $1,000 $600 **$300** Monthly cost $150,000 $30,000 $18,000 **$9,000** Annual cost $1.83M $365K $219K **$110K** Annual savings vs Opus — $1.46M $1.61M **$1.72M** * * * ## Architecture deep dives ### GLM-5 — Zhipu AI GLM-5 is trained **entirely on Huawei Ascend 910B chips** — achieving complete independence from NVIDIA hardware. At 744B total / 40B active, it has the highest active parameter count of the Chinese models. It uses DeepSeek Sparse Attention for long-context handling and was trained on 28.5 trillion tokens. Fully MIT-licensed. On Vals AI, GLM-5 scores highest among Chinese models on SWE-bench (67.80%) and Terminal-Bench (49.44%), suggesting strength in real-world software engineering tasks. Its AIME score (91.67%) is solid but trails Kimi K2.5 by 4 points. **Strengths:** Highest Vals Index among Chinese models (60.69%), strong SWE-bench, fully open-source, NVIDIA-independent. **Weaknesses:** Low IOI (22.00%), weaker GPQA (83.33%), most expensive Chinese option. ### MiniMax M2.5 The efficiency outlier. MiniMax M2.5 activates just **10 billion parameters** yet achieves 70.40% on SWE-bench — competitive with Qwen 3.5 and only 9 points behind Claude Opus 4.6, while activating a fraction of the parameters. The model uses 256 experts with 8 active per token across 230B total parameters. At $0.16 per test on Vals AI with 264s average latency, it is the cheapest and fastest model in the entire comparison. **Strengths:** Lowest cost ($0.16/test), fastest latency (264s), strong SWE-bench for its parameter budget. **Weaknesses:** Lowest Vals Index (53.57%), weak IOI (6.67%), trails on enterprise benchmarks. ### Kimi K2.5 — Moonshot AI The largest model at **1 trillion total parameters** (32B active, 384 experts with 8 active per token). Kimi K2.5 is natively multimodal — jointly pre-trained on approximately 15 trillion mixed text and vision tokens. Its **Agent Swarm** feature orchestrates up to 100 parallel sub-agents, trained via Parallel-Agent Reinforcement Learning (PARL). On Vals AI, it ties Claude Opus 4.6 on AIME (95.63%) and takes #1 on CorpFin (68.26%). Strong MMMU (84.34%) confirms its multimodal capability. **Strengths:** Best math among Chinese models (AIME: 95.63%), #1 CorpFin, strong MMMU, Agent Swarm parallelism. **Weaknesses:** SWE-bench (68.60%) trails GLM-5, Terminal-Bench (40.45%) is lowest in group. ### Qwen 3.5 — Alibaba The **context window champion** at 991K tokens — nearly 5x Claude Opus 4.6's 200K. At 397B total / 17B active, it uses a hybrid attention mechanism for efficient long-context handling. On Vals AI, Qwen 3.5 beats Claude Opus 4.6 on LiveCodeBench (85.33% vs 84.68%) and leads among Chinese models on GPQA (87.37%) and MMLU-Pro (87.18%). It ties MiniMax M2.5 on SWE-bench at 70.40%. **Strengths:** Massive context (991K), strongest GPQA and MMLU-Pro among Chinese models, leads LiveCodeBench. **Weaknesses:** Weakest AIME (86.04%), MMMU evaluation issues, higher latency (575s). * * * ## When to use which model **SWE-bench (real-world coding) → Claude Opus 4.6.** At 79.20%, it leads by 9+ points. No Chinese model is within striking distance on the standardized harness. **Competitive programming → Qwen 3.5.** At 85.33% on LiveCodeBench, it edges out even Claude Opus 4.6 (84.68%). Kimi K2.5 (83.87%) is also strong here. **Math → Kimi K2.5.** Ties Claude Opus 4.6 on AIME at 95.63% — at a fraction of the price. **Enterprise tasks → Claude Opus 4.6 or Kimi K2.5.** Opus leads most enterprise benchmarks. Kimi K2.5 takes #1 on CorpFin and is competitive on TaxEval. **Massive context → Qwen 3.5.** 991K tokens is unmatched. Particularly valuable for document processing, RAG, and multilingual workloads. **Cost optimization → MiniMax M2.5.** At $0.16/test and 264s latency, it is by far the cheapest and fastest option. At scale, this translates to $1.7M+/year in savings vs Claude Opus 4.6. **General-purpose frontier → Claude Opus 4.6.** Highest Vals Index (65.98%), leads SWE-bench, Terminal-Bench, most enterprise benchmarks. * * * ## The bigger picture The February 2026 model releases show Chinese labs closing the gap — but on standardized evaluation, Claude Opus 4.6 still leads on most benchmarks. The gap is narrowest on math (Kimi K2.5 ties Opus on AIME) and competitive programming (Qwen 3.5 edges Opus on LiveCodeBench). It is widest on software engineering (Opus leads SWE-bench by 9+ points) and terminal tasks (Opus leads by 9–18 points). For teams running high-throughput background agents, the question is: which task are you optimizing for? If it is a well-defined, repetitive task where Chinese models score within a few points of frontier — and it often is — the 5–17x cost savings are real. That is exactly what Maniac optimizes. We evaluate these models on your production data, fine-tune task-specialized variants, and route to the best quality-per-dollar model — so you get the right answer at a fraction of the cost. * * * _All benchmark scores sourced from [Vals AI](https://www.vals.ai) standardized evaluations, updated as of February 24, 2026. Vals AI uses the same evaluation harness for all models to ensure fair comparison. Pricing from official API documentation as of February 2026._ --- *Maniac — High throughput background agents. Opus-quality outputs at 1/50 of the cost. Learn more at [maniac.ai](https://www.maniac.ai).*