Three open-weight language models shipped inside an eight-week window this summer, and none of them came from OpenAI, Anthropic, or Google. Moonshot AI released Kimi K3 on July 16, 2026, with full weights following on July 27. Alibaba followed with Qwen3.8-Max on August 3, 2026, its first “Max-class” model ever released with open weights. Zhipu AI (operating internationally as Z.ai) had already shipped GLM-5.2 on June 17, 2026. Together they form the sharpest price-to-capability fight in open-weight AI right now, and the gap between the cheapest and most expensive of the three runs to $10.60 per million output tokens.
This comparison breaks down what each model actually costs, how they score on GPQA Diamond and SWE-bench, who is already running them in production, and which one makes sense for your workload. All three are open-weight, all three claim frontier-adjacent reasoning scores, and all three are priced far below GPT-5.6 or Claude Opus 5. The differences that matter come down to architecture, licensing terms, and what each lab optimized for.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Three Open-Weight Models Launched in Eight Weeks, What Changed
Open-weight AI stopped being a consolation prize sometime in the past year. Through most of 2025, teams that wanted GPT-4-class reasoning without a closed API had to accept a real capability tax. That tax has mostly disappeared. Kimi K3, Qwen3.8-Max, and GLM-5.2 all post GPQA Diamond scores above 91%, a graduate-level science reasoning benchmark that closed frontier models were leading almost exclusively a year ago.
What forced the timing is competitive pressure between Chinese AI labs racing to claim the open-weight throne before Western labs can respond with their own permissive releases. Moonshot AI, Alibaba, and Zhipu AI are all effectively subsidizing distribution: none of these companies makes its primary revenue from API tokens the way OpenAI or Anthropic does, so undercutting on price is a growth strategy, not a loss leader they need to fix later. That is the market condition that produced a $10.60-per-million-token output price gap between GLM-5.2 and Kimi K3, two models that trade blows on the same benchmark suite.
The practical effect for engineering teams: workloads that needed a Claude Opus 5 or GPT-5.6 subscription six months ago can now often run on a self-hosted or API-hosted open-weight model at a fraction of the cost, with a license that lets you fine-tune, redistribute, or run entirely on-premises. That is the trade this article is built to help you evaluate.
Kimi K3: Moonshot AI’s 2.8-Trillion-Parameter Bet
Moonshot AI, the Beijing-based lab behind the Kimi assistant, launched K3’s API on July 16, 2026, with its full open-weight rollout completing by July 27, 2026. The model is a sparse Mixture-of-Experts architecture at roughly 2.8 trillion total parameters, built from 896 experts with only 16 active per token according to architecture documentation published in July 2026, and positioned by Moonshot as built for coding, knowledge work, and multi-step reasoning tasks that run with minimal human supervision. That framing matters: Moonshot is explicitly targeting the agentic-workflow use case rather than chat-first consumer deployment.
K3 supports a 1-million-token context window, matching both competitors in this comparison. On independent benchmark aggregation, K3 currently posts the highest GPQA Diamond score of any open model tracked, at 93.5%, edging out both Qwen3.8-Max and GLM-5.2 on that specific graduate-level reasoning test. Moonshot has not published a full official benchmark card with SWE-bench or LiveCodeBench figures at the time of writing, so those numbers should be treated as pending rather than absent. The model ships as open weights, though the exact license text Moonshot is using has not been independently verified from a primary source as of this writing, only that download and self-hosting are permitted.
Pricing is where K3 stands apart, and not in its favor. As of August 2026, Moonshot’s API lists K3 at $3.00 per million input tokens and $15.00 per million output tokens, the highest of the three models by a wide margin and notably steeper than Qwen3.8-Max’s lower rate on the same rate-card snapshot, according to AIToolsReview’s pricing comparison. That premium buys the top GPQA score in the group, but teams running high-volume output workloads will feel it fast. You can check current published rates directly on Moonshot’s Kimi platform.
Qwen3.8-Max: Alibaba’s First Open-Sourced Max-Class Model
Alibaba shipped Qwen3.8-Max on August 3, 2026, and the release carries a distinction none of its predecessors did: it is the first model in Alibaba’s “Max” tier the company has ever open-sourced. Every prior Qwen Max release stayed behind Alibaba Cloud’s API. Weights for Qwen3.8-Max are distributed through Hugging Face and ModelScope, giving developers a genuine self-hosting path for a flagship-tier Qwen model for the first time. Alibaba kept iterating past launch, too: a refreshed Qwen3.8-Max-0902 snapshot shipped September 2, 2026, pushing its Code Arena WebDev score to 1,691 points, 17 points ahead of Kimi K3 Max’s 1,674, according to QwenCloud and DataCamp tracking.
The architecture is a 2.4-trillion-parameter Mixture-of-Experts model with 95 billion active parameters per token, which went generally available on August 3, 2026, with a 1-million-token context window, broken down as roughly 991,000 tokens of maximum input, 131,000 tokens of maximum output, and up to 262,000 tokens reserved for internal reasoning budget when the model runs in thinking mode. That reasoning-budget allocation is a distinct architectural choice from K3 and GLM-5.2, and it shows up in Qwen3.8-Max’s benchmark profile: 92.6% on GPQA Diamond, 67.7% on SWE-bench Pro, 73.5% on FrontierSWE, and 56.6% on DeepSWE 1.1.
Pricing sits in the middle of the pack: as of August 2026, Alibaba’s hosted API lists Qwen3.8-Max at $2.00 per million input tokens and $6.00 per million output tokens under its own custom Qwen3.8-Max licence, with cached input priced around $0.25 per million tokens for repeated context. Alibaba is also pushing Qwen3.8-Max into enterprise workflows directly through QwenWork, its workplace AI agent platform, positioning the model for repository-scale coding agents, long-document knowledge bases, long-video indexing, structured data extraction, and multi-step research assistants. Full specs and API access are published on Alibaba’s Qwen documentation site and the model repository on GitHub’s QwenLM organization.
GLM-5.2: Zhipu AI’s MIT-Licensed Contender
GLM-5.2 comes from Zhipu AI, the Chinese lab that operates its international product under the Z.ai brand. The model’s coding-plan tier launched June 13, 2026, with the broader public and open-weights release following on June 17, making GLM-5.2 the oldest of the three models in this comparison by roughly six weeks, and the most battle-tested in production as a result.
GLM-5.2 matches the 1-million-token context window of its rivals. On an independent open-source model leaderboard tracking July 2026 releases, GLM-5.2 ranked as the strongest open-source model that month, scoring 91.2% on GPQA Diamond. That places it just behind Qwen3.8-Max and Kimi K3 on that single benchmark, though GLM-5.2’s real advantage shows up elsewhere: cost. Multiple model-hosting platforms and secondary reports consistently describe GLM-5.2 as released under an MIT license, though Zhipu AI’s own primary licensing page was not directly verifiable at the time of this writing. If accurate, an MIT license would make GLM-5.2 the least restrictive of the three for commercial redistribution and fine-tuning.
Pricing is GLM-5.2’s clearest selling point: $1.40 per million input tokens and $4.40 per million output tokens, with cached input around $0.26 per million tokens, roughly matching Qwen3.8-Max’s cache rate but undercutting it on both base prices. NVIDIA’s own model documentation lists GLM-5.2 as suited for long-horizon reasoning, complex software engineering, agentic workflows, mathematical reasoning, coding and debugging, terminal automation, and general conversational use, a broad positioning that mirrors how enterprises are actually deploying it. You can review the model card directly through Z.ai’s platform or find hosted weights on Hugging Face.
Full Specs Comparison
Here is every core spec for all three models, lined up side by side. Numbers marked “not publicly disclosed” reflect gaps in each lab’s official documentation rather than omissions on our part, and we are not filling those in with estimates.
| Spec | Kimi K3 | Qwen3.8-Max | GLM-5.2 |
|---|---|---|---|
| Developer | Moonshot AI | Alibaba | Zhipu AI / Z.ai |
| Release date | July 16, 2026 (API); July 27, 2026 (weights) | August 3, 2026 | June 13, 2026 (coding plan); June 17, 2026 (public/open weights) |
| Architecture | Mixture-of-Experts | Mixture-of-Experts | Not publicly disclosed |
| Total parameters | ~2.8 trillion | ~2.4 trillion | Not publicly disclosed |
| Context window | 1,000,000 tokens | 1,000,000 tokens (991K input / 131K output / 262K reasoning) | 1,000,000 tokens |
| Open weights | Yes | Yes (first Max-class Qwen release) | Yes |
| License | Open-weight (exact terms unconfirmed) | Open-weight, distributed on Hugging Face and ModelScope | Reported MIT license (unverified from primary source) |
| GPQA Diamond | 93.5% | 92.6% | 91.2% |
| SWE-bench Pro | Not publicly disclosed | 67.7% | Not publicly disclosed |
| FrontierSWE | Not publicly disclosed | 73.5% | Not publicly disclosed |
| Input price / 1M tokens | $3.00 | $2.00 | $1.40 |
| Output price / 1M tokens | $15.00 | $6.00 | $4.40 |
| Cached input price / 1M tokens | Not publicly disclosed | ~$0.25 | ~$0.26 |
| Primary distribution | Kimi platform, OpenRouter | Alibaba Cloud, QwenWork, Hugging Face | Z.ai platform, Hugging Face, NVIDIA NIM |
The pattern across the table is consistent: Kimi K3 leads on GPQA Diamond but costs the most to run. Qwen3.8-Max was long the only one of the three with a publicly disclosed coding score, but Aireiter’s August 2026 benchmark comparison now credits Kimi K3 with an 81.2% FrontierSWE score, ahead of Qwen3.8-Max’s 73.5%, narrowing that gap. GLM-5.2 trails slightly on the single benchmark all three publish, but it is the cheapest model to operate by a comfortable margin and has been in production the longest.
Pricing Comparison: API Costs Per Million Tokens
Token pricing is the single biggest differentiator between these three models, since their reasoning benchmarks sit within about two points of each other. The table below shows list pricing plus a real-world cost projection for a workload of 10 million input tokens and 2 million output tokens, roughly what a mid-sized team running a coding assistant or document-processing pipeline might burn through in a week.
| Model | Input $/1M | Output $/1M | Cached input $/1M | Cost: 10M in + 2M out |
|---|---|---|---|---|
| Kimi K3 | $3.00 | $15.00 | Not disclosed | $60.00 |
| Qwen3.8-Max | $2.00 | $6.00 | ~$0.25 | $32.00 |
| GLM-5.2 | $1.40 | $4.40 | ~$0.26 | $22.80 |
Run that math out and GLM-5.2 costs 62% less than Kimi K3 for the same workload, and 29% less than Qwen3.8-Max. On output tokens specifically, GLM-5.2 undercuts Kimi K3 by $10.60 per million, a 70.7% discount. Against Qwen3.8-Max, GLM-5.2 is 26.7% cheaper on output. None of these prices include the operational cost of self-hosting, which changes the math entirely if you have GPU capacity sitting idle, but for API-hosted usage, the ranking is unambiguous: GLM-5.2 for cost, Kimi K3 for the top benchmark score, Qwen3.8-Max as the balance between the two.
Benchmark Performance Compared
GPQA Diamond, a 198-question graduate-level science benchmark covering physics, chemistry, and biology, is the one score all three labs have made public, and it comes from three separate reporting sources. Kimi K3’s 93.5% figure comes from an independent benchmark aggregator tracking open models rather than from Moonshot’s own marketing. Qwen3.8-Max’s 92.6% score appears consistently across Alibaba’s own announcement materials and third-party benchmark trackers including CometAPI’s model comparison tables. GLM-5.2’s 91.2% score comes from a July 2026 open-source model leaderboard that ranked it the top open model that month, a ranking later benchmarks from Kimi K3 and Qwen3.8-Max have since overtaken.
The spread across all three, 93.5% to 91.2%, is narrow enough that it should not be the deciding factor for most teams. A 2.3-point gap on a single benchmark rarely translates into a noticeable difference in day-to-day output quality, especially once you factor in prompt engineering and task-specific fine-tuning. Where the models diverge more meaningfully is on coding-specific evaluations: Qwen3.8-Max’s 67.7% SWE-bench Pro score remains the only published SWE-bench figure among the three, but FrontierSWE and Terminal-Bench 2.1 results published in August 2026 fill in more of the picture, Kimi K3 scores 81.2% on FrontierSWE against Qwen3.8-Max’s 73.5%, and on Terminal-Bench 2.1, Kimi K3 edges ahead again at 88.3% versus Qwen3.8-Max’s 86.6%, per Aireiter and AIToolsReview. That data gap is narrowing, but it is still worth weighing if code generation and repository-scale editing are your primary use case, since Qwen3.8-Max remains the only one of the three with a documented SWE-bench Pro track record specifically.
Independent tracking sites such as Artificial Analysis and the community leaderboard at SWE-bench.com periodically re-run standardized evaluations against newly released open-weight models, and both are worth checking directly for the latest third-party numbers before committing to a production deployment, since self-reported scores from any lab should be treated as a starting point rather than a final answer.
Benchmark methodology also varies enough between labs that raw score comparisons deserve some skepticism. GPQA Diamond is administered slightly differently across evaluation harnesses, some allow a scratchpad for intermediate reasoning, others cap the number of retry attempts, and few labs publish the exact prompt template they used to generate their reported score. A 93.5% versus a 91.2% result can reflect a genuine capability gap, or it can reflect two different test conditions. That is one more reason the pricing gap in this comparison is easier to trust than the benchmark gap: dollar figures do not depend on evaluation harness choices.
Coding and Agentic Task Performance
All three models are being marketed with agentic workflows as a primary use case, not an afterthought. Moonshot explicitly built K3 for “coding, knowledge work, and reasoning with minimal human supervision,” language that signals long-running autonomous tasks rather than single-turn chat completions. Qwen3.8-Max’s positioning leans the same direction, with Alibaba naming repository-scale coding agents and multi-step research assistants as flagship use cases. GLM-5.2’s NVIDIA documentation lists agentic workflows, terminal automation, and coding and debugging explicitly among its recommended applications.
In practice, that shared positioning means all three are reasonable choices for building a coding agent, but the evidence quality differs. Qwen3.8-Max has the only published SWE-bench Pro number in this group, which makes it the safer default if you need a defensible benchmark citation for a procurement decision. Kimi K3’s strength shows up anecdotally instead: reports place it inside the stack at companies including Cursor, which reportedly built parts of its product on a Kimi model foundation, a signal that the model performs well on real coding-agent workloads even without a published SWE-bench figure to point to. GLM-5.2’s terminal automation and debugging focus suggests it is tuned for DevOps-adjacent agent tasks specifically, distinct from the broader “write and edit code” framing of the other two.
If your agent needs to operate for extended, multi-step sessions without human checkpoints, the 262,000-token reasoning budget built into Qwen3.8-Max’s architecture is a genuine differentiator: it gives the model dedicated space to work through intermediate steps before producing a final answer, rather than consuming your visible context window with scratch work.
Context Window and Long-Document Handling
All three models converge on a 1-million-token context window, which by itself makes them roughly equivalent for feeding in large codebases, legal documents, or research corpora. The difference is in how that million tokens gets allocated. Qwen3.8-Max is the only model of the three with a published breakdown: about 991,000 tokens available for input, 131,000 for output, and up to 262,000 reserved for internal reasoning when thinking mode is active. Neither Moonshot nor Zhipu AI has published an equivalent breakdown for K3 or GLM-5.2, so it is not possible to say for certain how their million-token budgets split between input, output, and reasoning.
For long-document use cases specifically, Alibaba is marketing Qwen3.8-Max directly at long-document knowledge bases and long-video indexing, which suggests internal testing has validated performance at the far end of that context window rather than just the ability to accept the tokens. That is a meaningful distinction: plenty of models advertise large context windows while showing degraded recall past a few hundred thousand tokens, a well-documented pattern often called “lost in the middle.” Whether Kimi K3 and GLM-5.2 hold up as well at the top of their context windows is not something either lab has published data on yet, so teams with genuinely long-document workloads should run their own retrieval-accuracy tests before committing to either.
Real-World Adoption and Use Cases
Reported production adoption gives a useful signal beyond benchmark scores, since companies deploying a model at scale are implicitly vouching for its reliability under real traffic. Here is what has surfaced publicly for each model:
- DoorDash has reportedly adopted a Kimi model for internal, lower-priority tasks, reserving a different model for higher-stakes functions, a common pattern of routing cheaper or more specialized models to non-critical workloads.
- Coinbase has reportedly confirmed internal use of a Kimi model, though the exact scope of that deployment has not been publicly detailed.
- Cursor, the AI coding editor, reportedly built parts of its product on a Kimi model foundation, one of the stronger third-party validation signals for K3’s coding-agent performance.
- Alibaba Cloud’s QwenWork platform runs Qwen3.8-Max as its default workplace AI agent, putting the model directly in front of enterprise customers for document processing, research assistance, and coding tasks inside Alibaba’s own commercial ecosystem.
- NVIDIA has published GLM-5.2 as an available NIM microservice, recommending it for long-horizon reasoning, complex software engineering, agentic workflows, and terminal automation, an integration that gives enterprise NVIDIA customers a supported, GPU-optimized deployment path for the model.
Beyond named deployments, all three models are available through OpenRouter, which lets developers route requests across dozens of models through a single API and compare real-world latency and cost without committing to one vendor’s SDK. That kind of aggregator access is often the fastest way for a smaller team to test all three models against its own workload before making a hosting decision.
Other Open-Weight Contenders Worth Watching
Kimi K3, Qwen3.8-Max, and GLM-5.2 are the three most relevant open-weight releases for a head-to-head price and benchmark comparison right now, but they are not the entire field. DeepSeek V4 Pro left preview and became generally available on August 12, 2026, priced at $0.435 per million input tokens on a cache miss and $0.87 per million output tokens, undercutting all three models in this comparison on raw price while posting an 80.6% SWE-bench Verified score and a 1-million-token context window with a 384,000-token maximum output. DeepSeek V4-Flash followed on July 31, 2026, with a retrained pipeline aimed specifically at coding tasks. Both are worth a separate look if price is your only constraint, and tech-insider.org has covered that comparison directly in a dedicated GPT-5.6 vs DeepSeek V4 Pro breakdown.
MiniMax M3 is another open model worth flagging, appearing on July 2026 leaderboards with a GPQA Diamond score near 93% and a matching 1-million-token context window, putting it in the same performance tier as the three models compared here. Alibaba’s own Qwen3 235B-A22B, a smaller model than 3.8-Max, separately posted 85.7% on AIME 2024 and 70.7% on LiveCodeBench v5 on an independent open-source leaderboard, numbers worth checking if your workload is math- or competitive-programming-heavy rather than general reasoning. None of these unseat the three-way comparison at the center of this article, but they underline how crowded and fast-moving the open-weight field has become in the back half of 2026.
Licensing and Self-Hosting Differences
Open weights do not automatically mean open license, and the fine print matters if you plan to fine-tune, redistribute, or build a commercial product on top of any of these three models. GLM-5.2 is the model most consistently described as MIT-licensed across model-hosting platforms and secondary reporting, which would make it the least restrictive option for commercial use, redistribution, and derivative fine-tuning, if that reporting holds up against Zhipu AI’s own primary documentation. That caveat matters enough to repeat: verify the license directly against Zhipu AI’s official terms before building a commercial dependency on it, since a single independently-verified primary source was not available at the time of this research.
Qwen3.8-Max’s open-weight release is well documented as a first for Alibaba’s Max tier, with weights distributed through both Hugging Face and ModelScope, Alibaba’s own model hub, and Codersera confirms the full open-weight rollout for the 2.4-trillion-parameter, 95-billion-active-parameter model was complete as of August 12, 2026, marking the first time an Alibaba Qwen-Max model has shipped as open weights. Alibaba is licensing the release under its own custom Qwen3.8-Max licence rather than a fully permissive license like Apache 2.0 or MIT, so teams should check the specific terms attached to the 3.8-Max release rather than assuming continuity with smaller Qwen models that used more permissive terms.
Kimi K3’s open-weight status is confirmed, meaning the model can be downloaded and run on your own infrastructure, but Moonshot has not published clear, independently verified license terms as of this writing. Teams evaluating K3 for a commercial, self-hosted deployment should treat the license as an open question to resolve directly with Moonshot before committing engineering time to a self-hosted build.
Why Chinese Labs Are Driving This Price War
It is worth asking why three separate labs converged on aggressive open-weight pricing within the same eight-week stretch instead of holding capability behind a closed API the way OpenAI and Anthropic largely have. Part of the answer is business model. Moonshot AI, Alibaba, and Zhipu AI do not depend on token revenue as their core business the way a pure-play model company does. Alibaba sells cloud infrastructure, and a widely-adopted open Qwen model drives Alibaba Cloud compute consumption regardless of who pays for the tokens. Zhipu AI has built a broader enterprise software business around GLM, where the model itself functions as a distribution wedge. Moonshot is the closest of the three to a pure model company, and its higher K3 pricing reflects that difference in incentive.
Part of the answer is also geopolitical. US export controls on advanced AI chips have pushed Chinese labs toward efficiency-focused architectures and toward open distribution as a way to build global developer mindshare that does not depend on direct access to the largest US cloud markets. Anthropic’s Claude Fable 5 was itself briefly suspended in July 2026 under a US export-control directive before being restored, a reminder that regulatory friction now cuts both ways in this market. Open weights sidestep some of that friction entirely: once a model is downloadable, it runs wherever a customer’s hardware happens to be, regardless of which cloud region or API gateway sits in between.
Cost at Scale: What a Real Workload Costs
List pricing per million tokens is easy to compare but hard to feel until it is mapped onto an actual workload. Take a mid-sized SaaS company running an AI coding assistant that processes roughly 500 million input tokens and 100 million output tokens per month, a realistic figure for a product with a few thousand active daily users generating code completions and chat responses.
| Model | Monthly input cost (500M tokens) | Monthly output cost (100M tokens) | Total monthly cost |
|---|---|---|---|
| Kimi K3 | $1,500.00 | $1,500.00 | $3,000.00 |
| Qwen3.8-Max | $1,000.00 | $600.00 | $1,600.00 |
| GLM-5.2 | $700.00 | $440.00 | $1,140.00 |
At that volume, the gap between Kimi K3 and GLM-5.2 is $1,860 per month, or roughly $22,320 a year, for functionally similar reasoning capability. That is not a rounding error for a startup watching burn rate. It is the kind of number that should push any team seriously evaluating these models to run a real accuracy comparison on their own task set before defaulting to the highest-scoring benchmark model, since the practical cost of a 2.3-point GPQA gap rarely justifies a 163% price premium at scale.
Which Model Fits Which Use Case
No single model wins across every dimension, so the right pick depends heavily on what you are optimizing for. Here is how the three break down against common deployment scenarios.
- High-volume customer support or chat products: GLM-5.2. Output-heavy workloads amplify the per-token savings, and its $4.40 output price makes it the cheapest option for products generating long conversational responses at scale.
- Enterprise coding agents with procurement requirements: Qwen3.8-Max. It is the only model of the three with a published SWE-bench Pro score, giving you a defensible, citable benchmark for a purchasing decision, plus direct enterprise integration through QwenWork.
- Autonomous, long-running agent workflows: Kimi K3. Moonshot built it specifically for tasks that run with minimal human supervision, and its top GPQA Diamond score suggests an edge on complex, multi-step reasoning chains where accuracy compounds over many steps.
- Long-document RAG and knowledge-base search: Qwen3.8-Max, given Alibaba’s direct marketing toward long-document knowledge bases and the model’s published context-window breakdown.
- Data-residency-sensitive or air-gapped deployments: GLM-5.2 or Kimi K3, both available as downloadable weights for on-premises hosting, though confirm the license terms for your specific commercial use case before deployment.
- Startups testing multiple models before committing: Route all three through OpenRouter first. Comparing real task performance against your own prompts for a week costs far less than committing engineering time to a single vendor’s SDK and discovering a fit problem later.
Migration Guide: Moving From a Closed API to an Open-Weight Model
Switching a production workload from a closed model like GPT-5.6 or Claude Opus 5 to an open-weight alternative is not a one-line API key swap, even though the request format often looks similar. Here is a practical path for teams making that move.
- Benchmark your own task set first. Public benchmarks like GPQA Diamond and SWE-bench Pro are useful signals, but they rarely match your actual prompts. Pull 50 to 100 real production requests and run them against all three models before choosing one.
- Start with an aggregator, not a direct integration. Route initial testing through OpenRouter so you can A/B test Kimi K3, Qwen3.8-Max, and GLM-5.2 against your current model without three separate SDK integrations.
- Audit your prompts for provider-specific formatting. System prompt structure, function-calling syntax, and stop-token handling differ between providers. Expect to adjust prompts, not just swap endpoints.
- Test context window behavior at your actual document lengths, not just at the advertised 1-million-token maximum. Retrieval accuracy often degrades well before the stated limit.
- Verify the license against your commercial use case. Confirm redistribution, fine-tuning, and commercial-use terms directly with the vendor before building a dependency, particularly for Kimi K3 given its currently unconfirmed license text.
- Run a shadow deployment before cutting over. Send production traffic to the new model in parallel with your existing one, log both outputs, and compare quality on a sample before routing real users to it.
- Model your cost at your actual monthly volume, not list price per million tokens. Use the workload tables above as a template and plug in your real input/output token counts.
- Plan for a rollback path. Keep your existing provider integration live and toggled by a feature flag for at least one full billing cycle after cutover, in case the new model underperforms at scale in ways your shadow test missed.
Teams that have already gone through a similar migration for DeepSeek V4 Pro can reuse most of this checklist, since the underlying decision points, cost modeling, prompt auditing, and license verification, carry over regardless of which open-weight model you land on.
Pros and Cons of Each Model
Kimi K3
Pros: Highest GPQA Diamond score in the group at 93.5%. Built specifically for long-running, minimally-supervised agentic tasks. Reported adoption at Cursor, DoorDash, and Coinbase suggests solid real-world coding-agent performance.
Cons: Most expensive of the three by a wide margin at $15.00 per million output tokens. No published SWE-bench or LiveCodeBench score. License terms not independently confirmed as of this writing.
Qwen3.8-Max
Pros: Only model of the three with a published SWE-bench Pro score (67.7%). Detailed, transparent context-window breakdown. Direct enterprise integration through QwenWork and Alibaba Cloud. First open-sourced Max-class Qwen model, a meaningful precedent for future releases.
Cons: Mid-tier pricing, not the cheapest option. Slightly behind Kimi K3 on GPQA Diamond. License terms for the 3.8-Max tier specifically have not been confirmed as fully permissive.
GLM-5.2
Pros: Cheapest of the three on both input and output pricing. Longest production track record, live since mid-June 2026. Reported MIT license would make it the most flexible for commercial redistribution. Native NVIDIA NIM support for GPU-optimized enterprise deployment.
Cons: Lowest GPQA Diamond score of the three, though the gap is under two points. No published SWE-bench score for coding-specific procurement decisions. MIT license status not independently confirmed from Zhipu AI’s own primary documentation.
The Verdict
There is no single winner here, and treating this as a “best model” ranking would misrepresent the actual trade-off. On price, GLM-5.2 wins clearly: $22.80 versus $60.00 for an identical 10-million-input, 2-million-output workload, a 62% savings over Kimi K3 for a GPQA score within two points of the top result. For teams running high-volume, cost-sensitive workloads, that math is close to decisive.
On raw reasoning benchmark performance, Kimi K3 leads, and its positioning toward autonomous, minimally-supervised agent tasks combined with reported adoption at Cursor gives it real-world credibility beyond the benchmark number. Teams building agent products where accuracy on hard, multi-step reasoning tasks matters more than per-token cost should weigh K3 seriously despite the price premium.
Qwen3.8-Max is the pragmatic middle choice, and arguably the safest default for teams that need a defensible, benchmarked answer for coding-agent procurement. It is the only model here with a published SWE-bench Pro score, it comes from a lab with deep enterprise distribution through Alibaba Cloud, and its pricing sits close enough to GLM-5.2 that the cost difference rarely decides the outcome on its own.
The practical recommendation: start with GLM-5.2 or Qwen3.8-Max for cost-sensitive or coding-focused production workloads, and reserve Kimi K3 for tasks where its top GPQA score and agentic-task focus justify the premium. Whichever you choose, run your own benchmark against your actual prompts before committing, since a 2.3-point spread on a single public benchmark is a weak substitute for testing on your real workload.
Frequently Asked Questions
Which model is cheapest: Kimi K3, Qwen3.8-Max, or GLM-5.2?
GLM-5.2 is the cheapest of the three, priced at $1.40 per million input tokens and $4.40 per million output tokens. Qwen3.8-Max costs $2.00/$6.00 per million tokens, and Kimi K3 is the most expensive at $3.00/$15.00 per million tokens.
Are Kimi K3, Qwen3.8-Max, and GLM-5.2 open source?
All three ship as open-weight models, meaning you can download and self-host them. Their exact license terms differ: GLM-5.2 is widely reported as MIT-licensed, Qwen3.8-Max is distributed through Hugging Face and ModelScope under Alibaba’s own terms, and Kimi K3’s precise license text had not been independently confirmed at the time of this writing. Verify the specific license against each vendor’s official documentation before commercial use.
Which model has the best benchmark scores?
Kimi K3 leads on GPQA Diamond at 93.5%, ahead of Qwen3.8-Max at 92.6% and GLM-5.2 at 91.2%. Qwen3.8-Max is the only model of the three with a published SWE-bench Pro score, at 67.7%, so it is the strongest documented choice if coding benchmarks specifically are your deciding factor.
Can I run these models on OpenRouter?
Yes, all three are accessible through OpenRouter, which lets you route requests to any of them through a single API and compare cost and latency without integrating each vendor’s SDK separately.
Which companies are using these models in production?
Reported adopters include DoorDash and Coinbase for internal Kimi model use, and Cursor for parts of its coding-editor product built on a Kimi model foundation. Qwen3.8-Max powers Alibaba Cloud’s QwenWork platform, and GLM-5.2 is available as a supported NVIDIA NIM microservice for enterprise deployment.
How big is the context window on each model?
All three support a 1-million-token context window. Qwen3.8-Max is the only one with a published breakdown: roughly 991,000 tokens of input, 131,000 of output, and up to 262,000 for internal reasoning. Kimi K3 and GLM-5.2 have not published an equivalent breakdown.
Should I switch from GPT-5.6 or Claude Opus 5 to one of these models?
It depends on your workload. All three open-weight models are priced well below proprietary frontier APIs and score within striking distance on public reasoning benchmarks, but closed models still lead on some specialized tasks and offer more mature tooling ecosystems. Run a shadow test against your real production traffic before committing to a full migration, following the checklist above.
Is GLM-5.2 or Qwen3.8-Max better for enterprise deployment?
Both have real enterprise distribution paths. Qwen3.8-Max integrates directly into Alibaba Cloud’s QwenWork platform, while GLM-5.2 is available as a supported NVIDIA NIM microservice, giving it a more GPU-vendor-neutral enterprise deployment story. Choose based on your existing cloud infrastructure and whether a defensible SWE-bench score matters for your procurement process.
Related Coverage
- Claude Opus 5 vs GPT-5.6 vs DeepSeek V4-Pro: $22 Gap [2026]
- Claude Sonnet 5 vs GPT-5.6 vs Gemini 3.7 Flash: 6.7x Price Gap [2026]
- Claude Fable 5 vs Grok 4.6 vs DeepSeek V4: 178x Price Gap [2026]
- GPT-5.6 vs DeepSeek V4 Pro 0813: 714x Cheaper Input [2026]
- How to Set Up OpenRouter: 13 Steps, 80 Min [2026]
- Best Open Source LLM [2026]: DeepSeek V4 Hits 80.6%
