Something strange happened to the AI industry over the last eight weeks.
In June, Z.ai shipped GLM-5.2 and quietly became the first open-weight model to cross 80% on Terminal-Bench. In July, Moonshot AI dropped Kimi K3 — 2.8 trillion parameters, the largest open-weight model ever published — and Chinese AI stocks fell off a cliff in sympathy. Then on August 3rd, Alibaba answered with Qwen3.8-Max, 2.4 trillion parameters, and open-sourced a Max-tier model for the first time in the company’s history.
Three frontier-adjacent models. Three different licensing philosophies. Three wildly different price points. And an enormous amount of confused coverage treating them as interchangeable.
They are not interchangeable. One of them can’t see images. One of them isn’t actually downloadable yet, despite a hundred headlines calling it “open source.” And one of them costs five times more per output token than another while scoring worse on several coding benchmarks.
This is the comparison that sorts it out.
TL;DR — The 30-Second Verdict
If you want the strongest raw capability and can afford it: Kimi K3. It ranks #4 on the Artificial Analysis Intelligence Index and #1 on Frontend Code Arena, and it’s the only one of the three that trades punches with Claude and GPT flagships. You’ll pay $3/$15 per million tokens for the privilege.
If you want the best cost-to-capability ratio for coding: GLM-5.2. At $1.40/$4.40 it’s less than a third of K3’s output price, it’s genuinely MIT-licensed, and the weights have been on Hugging Face since June. It just can’t process images.
If you need multimodal input and long-horizon agents in one model: Qwen3.8-Max. Text, image and video input, plus the strongest published vision-agent numbers of the three. But as of this writing the weights are still a promise, not a download.
If you’re building anything serious: Use more than one. The routing section below explains why that’s not a cop-out answer.
Table of Contents
- Why This Comparison Matters Right Now
- Meet the Contenders
- Head-to-Head Specifications
- Benchmark Deep Dive
- The Benchmark Asterisk Nobody Mentions
- Pricing: The Real Cost Math
- What “Open Weights” Actually Means Here
- Can You Actually Self-Host These?
- Three 1M Context Windows Are Not Equal
- Which Model Should You Pick?
- The Multi-Model Routing Strategy
- Risks, Caveats and Red Flags
- What’s Coming Next
- Frequently Asked Questions
- Final Verdict
Why This Comparison Matters Right Now
For most of 2024 and 2025, the open-weight conversation was a consolation bracket. You picked a Western frontier model for the hard work and an open model for the cheap, high-volume stuff. The gap was real and everybody knew it.
That framing broke this summer.
When Moonshot released Kimi K3 in mid-July, the market reaction told the story better than any benchmark table. Z.ai’s Hong Kong-listed shares dropped as much as 30%. MiniMax fell as much as 16%. Alibaba slid 4%. Investors weren’t reacting to a research paper — they were reacting to a model that developers, in blind side-by-side evaluations, were picking over leading American systems on front-end coding work.
Then Alibaba’s Qwen3.8-Max launch sent its own stock up 7% in Hong Kong. The same week, Moonshot’s daily revenue reportedly grew by at least a factor of six post-K3, with the company reaching $300 million in annual recurring revenue in June — up from $200 million in April — while seeking a funding round at a $50 billion valuation.
This is no longer a story about cheap alternatives. It’s a story about whether the price of frontier-class capability just collapsed, and if so, which of these three you should actually wire into production.
Meet the Contenders
Kimi K3 (Moonshot AI)
Released: July 16, 2026 (API) · July 27, 2026 (weights)
Kimi K3 is the scale play. At 2.8 trillion total parameters it’s roughly 75% larger than DeepSeek’s V4 Pro, which had held the “largest widely-used open model” title. But the number that matters for anyone thinking about inference cost is the sparse one: K3 uses a mixture-of-experts design with 896 experts and only 16 active per token, putting roughly 50 billion parameters to work on any given forward pass.
Two architectural pieces are Moonshot’s own: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, which the company describes as a drop-in replacement for standard residual connections that keeps paying off as you scale. Both had been published as open research on GitHub before K3 shipped.
The positioning is unambiguous — Moonshot built this for long-horizon engineering. The company’s framing is about sustaining multi-hour sessions, navigating large repositories and orchestrating terminal tools with minimal supervision. It ships with native visual understanding and an always-on reasoning mode, and it’s compatible with the OpenAI SDK, so integration is mostly a base-URL swap.
Qwen3.8-Max (Alibaba)
Released: July 19, 2026 (preview at WAIC Shanghai) · August 3, 2026 (general availability)
Qwen3.8-Max is 2.4 trillion total parameters with roughly 95 billion active per token — nearly double K3’s active count, which has real implications for serving cost that we’ll come back to.
The genuinely new thing here isn’t scale, it’s modality. This is the first Qwen model above a trillion parameters to go multimodal, accepting text, images and video as input. And it’s the first time Alibaba has committed to open-sourcing a Max-tier flagship at all. Every previous Max model — Qwen3.7-Max, Qwen3.6-Max-Preview — stayed locked behind the API while the open-weight line continued separately at smaller sizes.
Alibaba’s long-horizon demos are the most striking marketing of the three. In one, the model reportedly spent over ten days autonomously building a self-evolving software harness, incorporating feedback and iterating through code, previews and logs without human intervention. In another, it reproduced a machine learning research paper from scratch: 33 rounds of GPU training across roughly 125 hours, 7,600 lines of code, then 18 self-generated improvement ideas that beat the original paper’s method. Entered into a live contest against 526 human teams, it placed ahead of 87% of the field.
Treat those as vendor demos, because that’s what they are. But they’re unusually specific vendor demos.
GLM-5.2 (Z.ai / Zhipu AI)
Released: June 13, 2026 (Coding Plan preview) · mid-June 2026 (full open release)
GLM-5.2 is the smallest model here by a wide margin and the most sharply focused. Roughly 744–753 billion total parameters with about 40 billion active per token, it was built for one thing: long-horizon agentic coding.
That focus shows up as a deliberate omission. GLM-5.2 has no visual encoder. It accepts text and code. Feed it a screenshot, a PDF or a UI mockup and it errors out. Z.ai co-founder Jie Tang has been public about the reasoning — he views text-based reasoning, not visual input, as the capability that raises the ceiling on machine intelligence. A June community poll on next-version features drew over 466,000 views, and vision was overwhelmingly the top request. It still hasn’t shipped.
What GLM-5.2 does have is the cleanest licensing story in this comparison and, by some distance, the lowest price. It went out under an unrestricted MIT license with weights live on Hugging Face, and it arrived with day-one support across more than twenty third-party coding environments.
Head-to-Head Specifications
| Kimi K3 | Qwen3.8-Max | GLM-5.2 | |
|---|---|---|---|
| Developer | Moonshot AI | Alibaba | Z.ai (Zhipu AI) |
| Total parameters | 2.8T | 2.4T | ~744–753B |
| Active per token | ~50B (16 of 896 experts) | ~95B | ~40B |
| Architecture | Sparse MoE + Kimi Delta Attention | Sparse MoE | Sparse MoE + IndexShare attention |
| Context window | 1M tokens | 1M tokens (991K max input) | 1M tokens |
| Max output | — | 131K tokens | 128K–131K tokens |
| Text input | ✅ | ✅ | ✅ |
| Image input | ✅ | ✅ | ❌ |
| Video input | ❌ | ✅ | ❌ |
| Reasoning modes | Always-on thinking | Standard / thinking (262K budget) | High / Max |
| Weights available | ✅ July 27, 2026 | ⏳ Promised week of Aug 10 | ✅ Since June 2026 |
| License | Modified MIT (with conditions) | Not yet announced | MIT (unrestricted) |
| API input / 1M | $3.00 | $2.00 | $1.40 |
| API output / 1M | $15.00 | $6.00 | $4.40 |
| Cached input / 1M | $0.30 | $0.25 | $0.26 |
Three things jump out of that table.
First, GLM-5.2 is doing more with less. It’s roughly a quarter the total size of K3 and it’s competing on coding benchmarks — which tells you scale isn’t the whole story in 2026.
Second, Qwen3.8-Max’s active parameter count is the outlier. At ~95B active it’s nearly twice K3’s, despite being smaller in total. Active parameters drive serving cost, so anyone planning to self-host should note that Qwen’s flagship is more expensive to run per token than a model with 400 billion more parameters.
Third, the price spread is enormous. K3’s output tokens cost 3.4× GLM-5.2’s. Over a month of heavy agent usage, that is not a rounding error.
Benchmark Deep Dive
Coding and agentic performance
This is where all three models are aiming, so it’s where the comparison actually matters.
| Benchmark | Kimi K3 | Qwen3.8-Max | GLM-5.2 |
|---|---|---|---|
| Terminal-Bench 2.1 | 88.3 | 86.6 | 81.0 |
| SWE-bench Pro | — | 67.7 | 62.1 |
| FrontierSWE | 81.2 | 73.5 | 74.4 |
| SWE Marathon | 42.0 (#1) | — | — |
| Program Bench | 77.8 (#1) | — | — |
| DeepSWE | 67.5 | 56.6 | — |
| MCP-Atlas (tool use) | ~Fable 5 −0.5 | — | 77.0 |
For context on those Terminal-Bench numbers: GPT-5.6 Sol at maximum reasoning scores 88.8, and Claude Opus 4.8 and Claude Fable 5 both land at 84.6 in Alibaba’s reported table. Kimi K3’s 88.3 puts it half a point off the top. Qwen3.8-Max’s 86.6 clears both Claude models in that same table. GLM-5.2’s 81.0 was, at the time, the first open-weight score to cross 80.
GLM-5.2’s headline win is against GPT-5.5 specifically: 62.1 vs 58.6 on SWE-bench Pro, and 74.4 vs 72.6 on FrontierSWE — the latter putting it in a near-tie with Claude Opus 4.8 at 75.1. On MCP-Atlas tool usage it hit 77.0 against GPT-5.5’s 75.3 and Opus 4.8’s 77.8.
Reasoning and general knowledge
| Benchmark | Kimi K3 | Qwen3.8-Max | GLM-5.2 |
|---|---|---|---|
| GPQA Diamond | 93.5 | 92.6 | 89.5 |
| Artificial Analysis Intelligence Index | 57.11 (#4 of 189) | — | 51.09 |
| BrowseComp | 91.2 (#1) | — | — |
| PaperBench | — | 93.0 (leads) | — |
| IFBench | — | 82.8 (leads) | — |
K3’s 93.5 on GPQA Diamond was, at launch, the best score any open-weight model had published on that benchmark. On the independent Artificial Analysis Intelligence Index it scored 57.11, placing fourth of 189 tracked models — behind Claude Fable 5 (59.86) and both GPT-5.6 Sol configurations (58.89 and 57.65), but ahead of Claude Opus 4.8 at 55.69.
GLM-5.2 sits noticeably lower on that composite index at 51.09. That gap is the honest cost of its price advantage on general reasoning — though it narrows sharply on pure coding work.
Multimodal and vision
This is the category where GLM-5.2 simply doesn’t compete, and where Qwen3.8-Max makes its strongest case.
| Benchmark | Qwen3.8-Max | Notes |
|---|---|---|
| OSWorld-Verified | 86.1 | Computer-use agent tasks |
| Parametric CAD Bench | 91.5 | Structured design reasoning |
| OmniDocBench 1.5 | 92.1 | Document understanding |
| Vision Arena | #2 globally | Behind only a Fable 5 variant |
Qwen3.8-Max tops most vision rows in Alibaba’s published table, and its second-place Vision Arena ranking is a genuinely strong independent signal. Kimi K3 has native visual understanding and reportedly performs well on chart, document and multimodal work, but Qwen owns this category among the three.
Human preference rankings
Blind, crowdsourced arena results are worth reading separately from benchmark tables, because they capture something the tables miss.
- Frontend Code Arena: Kimi K3 ranked #1 with 1,679 Elo, ahead of Claude Fable 5 at 1,631. Qwen3.8-Max ranked #4 with 1,668 — 37 points behind Claude Opus 5.
- Text Arena: Qwen3.8-Max entered as the highest-ranking Chinese model, though still trailing several Anthropic offerings.
- Code Arena: GLM-5.2 took second place among coding models on the blind leaderboard.
Kimi K3 winning Frontend Code Arena outright — over Fable 5 — is arguably the single most impressive result in this entire article.
The Benchmark Asterisk Nobody Mentions
Every table above deserves a health warning, and the honest version of this comparison has to say so plainly.
Harness effects are enormous. A coding-agent benchmark doesn’t measure a model. It measures a model plus the runtime around it: tool access, file permissions, context compaction logic, retry behaviour, stopping rules. Kimi Code, Claude Code, Codex, Terminus and mini-SWE-agent are all different harnesses with different capabilities. A result labelled “Kimi K3 + Kimi Code” is not directly comparable to a bare API result.
Moonshot’s own launch table mixes harnesses — KimiCode and Claude Code for K3, Claude Code for the Claude and GLM rows, Codex for GPT. Some competitor scores come from those competitors’ own vendors. In some evaluations, Fable 5 requests rejected under usage policy fell back to Opus 4.8. There’s a useful sanity check buried in the data, though: K3 scored 67.5 on DeepSWE with KimiCode and 67.3 with the benchmark’s own mini-SWE-agent harness. A 0.2-point spread suggests the harness effect on that particular task is small.
Vendor tables flatter their authors. Alibaba’s multimodal comparison benchmarks Qwen3.8-Max against Qwen3.7-Plus, not Qwen3.7-Max — which makes the generational jump look larger than it is. To Alibaba’s credit, the company also published an RL scaling curve that doesn’t flatter it: performance peaks at 0.725 around 4,000 training environments, then declines to 0.719 and 0.689. That’s a company showing its diminishing returns in public, which is more than most labs do.
Independent verification lags. GLM-5.2 shipped with no benchmark table at all; the numbers circulating came later from vendor data and third-party testing. Qwen3.8-Max’s five published scores are vendor-reported, with no independent evaluation available as of early August. Kimi K3 has the most third-party data behind it — Artificial Analysis, Vals, Arena — and even there the results disagree with each other. Vals ranked K3 second overall on July 17, between Fable 5 and GPT-5.6 Sol, while placing Sol ahead of K3 specifically on Terminal-Bench 2.1 and SWE-bench Verified.
The most useful independent test I found came from Fireworks AI, which ran K3 against Claude Fable 5 across roughly 1,000 agentic tasks. The result wasn’t a clean win for anyone. K3 came out ahead on security work, crypto and long terminal loops. Fable 5 won on multilingual tasks and web/data visualisation. Neither dominated.
The practical takeaway: run your own evaluation on your own tasks before committing. Every one of these models is close enough to the others that leaderboard position won’t predict your results.
Pricing: The Real Cost Math
Per-token API rates
| Model | Input / 1M | Output / 1M | Cached input / 1M |
|---|---|---|---|
| GLM-5.2 | $1.40 | $4.40 | $0.26 |
| Qwen3.8-Max | $2.00 | $6.00 | $0.25 |
| Kimi K3 | $3.00 | $15.00 | $0.30 |
| Claude Sonnet 5 | $2.00 | $10.00 | — |
| Claude Opus 5 | $5.00 | $25.00 | $0.50 |
| Claude Fable 5 | $10.00 | $50.00 | — |
Two observations that matter more than the raw numbers.
Kimi K3 is priced like a Western mid-tier model, not a budget alternative. At $3/$15 it costs more per output token than Claude Sonnet 5 and 3.4× more than GLM-5.2. If you assumed “Chinese open-weight model” meant “cheap,” K3 breaks that assumption.
Caching is the real lever, not sticker price. Qwen’s cached input at $0.25 is eight times cheaper than fresh input, and Alibaba’s own documentation makes the point that prefix stability drives cost more than prompt length does. For agent workloads with a stable system prompt and a growing conversation, aggressive caching will move your bill more than model choice will. Qwen also offers explicit cache creation at $2.50 and explicit reads at $0.17 for workloads you can plan around.
Subscription plans
If you’re a solo developer or small team, per-token pricing may not be how you buy at all.
GLM Coding Plan has three tiers. Base monthly rates are $18 (Lite), $72 (Pro) and $160 (Max), with discounts by billing cycle — 10% monthly, 20% quarterly, 30% yearly. On annual billing that works out to roughly $12.60, $50.40 and $112 per month. Each tier gives you a prompt pool that refreshes every five hours, and the plan works out of the box with Claude Code, Cline, Kilo Code and OpenClaw.
One catch worth understanding: quota consumption uses multipliers. Off-peak GLM-5.2 usage sits at 1× through September 2026 as a promotion. After that it reverts to 2× off-peak and 3× peak — meaning your effective quota depends heavily on when you work, not just how much.
Qwen offers a Token Plan subscription starting around $6/month (39 CNY) at the Lite tier as an alternative to per-token billing.
Moonshot runs subscription tiers for the Kimi apps named after tempo markings — Moderato, Allegro, Allegretto and Vivace — with the apps themselves free under general rate limits.
Cost per task vs cost per token
Here’s the trap: per-token price is not cost per completed task.
A model that solves your problem in one pass at $15/1M output can easily be cheaper than one that needs four iterations at $4.40/1M. Reasoning tokens bill at output rates across all three of these models, and K3 in particular runs always-on thinking — one analysis noted K3 completed the Artificial Analysis suite at slightly lower total cost than GPT-5.6 Sol despite generating substantially more output tokens, because it needed fewer attempts.
Measure cost per resolved ticket, not cost per million tokens. It’s the only number that reflects reality.
What “Open Weights” Actually Means Here
This is the section where most coverage of these three models falls apart, because “open” is doing enormous work in a lot of headlines.
Three different states are being collapsed into one word:
- Available via API — you can call it, you can’t have it
- Open weights — you can download the parameters
- Open source with a permissive license — you can download, modify, redistribute and commercialise freely
Here’s where each model actually sits.
GLM-5.2: genuinely open
MIT license, unrestricted, weights live on Hugging Face under zai-org/GLM-5.2. Download, fine-tune, self-host, sell products built on it. No revenue thresholds, no attribution requirements, no negotiation. This is as clean as open-weight licensing gets, and it’s a real competitive advantage that gets under-discussed because MIT licensing isn’t a headline.
Kimi K3: open with conditions
Weights released July 27 under a custom license Moonshot calls the Kimi K3 License. Broad commercial use is permitted, but two conditions apply:
- Any company with annual revenue above $20 million must negotiate a contract with Moonshot before offering K3 to external customers as a service
- Any company with monthly revenue above $20 million or more than 100 million monthly active users must display Kimi K3 attribution in any product incorporating the model
For an indie developer or a startup, neither condition binds. For an established SaaS company planning to resell K3-powered inference, the first one absolutely does. Read the license before you build a business model on it.
Qwen3.8-Max: not yet open
This is the part that needs saying bluntly, because a great many headlines have already called Qwen3.8-Max open source.
As of August 5, 2026, Qwen3.8-Max weights are not downloadable. Alibaba committed at launch to releasing weights “next week” — meaning the week of August 10 — on Hugging Face and ModelScope, for both Qwen3.8-Max and a smaller Qwen3.8-27B checkpoint. No license has been named. A Hugging Face search for Qwen3.8 as of August 4 returned no official Alibaba model card, only community uploads with names like Qwen3.8_4B_Distilled — third-party distills, not the flagship, and easy to mistake for the real thing if you’re scanning quickly.
There’s reason for cautious optimism: Alibaba has a genuine open-source record, with the Qwen3 and Qwen3.5 lines shipping under Apache 2.0. But there’s also a pattern in the other direction — every Max-tier flagship since Qwen3-Max has been proprietary. If the weights land with a permissive license, it breaks that pattern entirely and is arguably a bigger story than the parameter count. Until a repository and a license file exist, plan around the API.
Can You Actually Self-Host These?
Open weights are only useful if you can run them, and this is where the marketing and the hardware reality diverge sharply.
Kimi K3 is a 1.56 TB download, shipped in MXFP4 precision for broader hardware compatibility. Even so, a 2.8-trillion-parameter model needs supernode-class infrastructure — 64 or more accelerators is the recommended figure. For the overwhelming majority of teams, the weights are a research artefact and a sovereignty guarantee, not a deployment plan. You will use the API.
Qwen3.8-Max at 2.4T total is likewise a multi-node datacentre artefact. Alibaba’s smaller Qwen3.8-27B checkpoint is the realistic on-premise path, and it’s the one that should interest anyone actually planning to self-host.
GLM-5.2 at ~753B total and ~40B active is the most tractable of the three, but “most tractable” still means serious GPU infrastructure. Small businesses generally won’t have it.
So why does open weighting matter if almost nobody can run these? Two reasons, both real.
The first is sovereignty. Nations spending billions on domestic AI infrastructure have mostly owned the hardware while leasing the models from American providers. An open-weight frontier model changes that calculus — a government can run it on its own soil, retrain it for its own language and legal context, and keep citizen data inside its borders.
The second is the inference market. Open weights mean any hosting provider can serve the model, which drives per-token prices down through competition rather than vendor goodwill. That’s a large part of why GLM-5.2 is available at $1.00–$1.20 input via third-party routers, below Z.ai’s own $1.40.
Three 1M Context Windows Are Not Equal
All three models advertise a 1-million-token context window. That’s roughly 750,000 words — an entire large codebase, or about 400 pages of documents, in a single request.
The details differ in ways that matter operationally:
- Qwen3.8-Max caps actual input at 991K tokens, dropping to 983K when thinking is enabled, with a separate 262K maximum reasoning budget. Output tops out at 131K tokens in both modes. Rate limits are 2M tokens per minute and 15K requests per minute.
- GLM-5.2 supports up to 128–131K output tokens, enough for large multi-file diffs in one shot. Its 1M window was a roughly 5× jump over GLM-5.1’s ~200K, and it’s the upgrade that most changed what the model can do.
- Kimi K3 offers 1M context with the strongest long-context retrieval signals of the three, given its architecture was designed around long-horizon sessions specifically.
The more useful question than “how big is the window” is “does the model stay coherent at the far end of it.” Advertised context and usable context diverge, and none of these three has been independently stress-tested at full length in a way I’d consider conclusive. If your workload depends on genuine 800K-token recall, test it before you architect around it.
Which Model Should You Pick?
For the solo developer or indie hacker → GLM-5.2
The economics are decisive. A Lite Coding Plan at $12.60/month on annual billing gets you a genuinely capable coding agent inside Claude Code or Cline, with a five-day free trial available through ZCode (3 million GLM-5.2 tokens plus 2 million GLM-5-Turbo tokens per day) to form a real judgement before paying anything.
You give up vision. If your work is backend, CLI, scripting or API development, you’ll rarely notice. If you’re doing front-end work from design mockups, you will notice immediately.
For the startup engineering team → GLM-5.2 primary, Kimi K3 for escalation
Route the wide middle of your work — routine bug fixes, refactors, test generation, boilerplate — to GLM-5.2 and reserve K3 for the tasks where the extra capability actually converts into a completed ticket. The 7-point SWE-bench Pro gap between them is real, but it only matters on the hard tail.
For enterprise and regulated industries → self-hosted GLM-5.2, or Qwen3.8-27B when it lands
The critical detail: Z.ai’s cloud API routes through China-based infrastructure. For personal projects and most startups this is a non-issue. For a regulated enterprise it demands either a legal review or a shift to self-hosting. The MIT license makes self-hosting genuinely viable in a way K3’s revenue-threshold conditions do not.
For multimodal and computer-use workflows → Qwen3.8-Max
If your agent needs to read screenshots, parse documents visually, or process video, GLM-5.2 is disqualified and Qwen leads Kimi on the published vision numbers. OSWorld-Verified at 86.1 and OmniDocBench at 92.1 are strong computer-use and document-understanding signals. Just budget for API access rather than self-hosting until the weights and license actually appear.
For researchers and sovereignty-focused deployments → Kimi K3
The largest open-weight checkpoint ever published, with a technical report covering design, training and benchmark methodology. If you’re studying frontier-scale MoE architectures, Kimi Delta Attention or Attention Residuals, this is the most capable model you can actually inspect.
For maximum capability regardless of cost → Kimi K3
Fourth on Artificial Analysis, second on the Vals Index, first on Frontend Code Arena, and the only model of the three that consistently trades wins with Claude and GPT flagships. It costs accordingly.
The Multi-Model Routing Strategy
The honest answer to “which one should I use” is that picking one is usually the wrong frame.
All three of these models are OpenAI-SDK compatible. GLM-5.2 additionally ships Anthropic-compatible endpoints, and Qwen supports both OpenAI and DashScope specs. Switching between them is a base URL and a model ID. There is almost no lock-in cost.
That makes a routing architecture cheap to build and expensive to skip:
- Cheap tier for high-volume, low-stakes work. Classification, summarisation, straightforward edits. GLM-4.7 at $0.60/$2.20 handles more than people expect — it still posts 73.8% on SWE-bench Verified.
- Default tier for the bulk of real work. GLM-5.2 or Qwen3.8-Max depending on whether you need vision.
- Escalation tier for the hard tail. Kimi K3, or a Western flagship, for tasks where failure is expensive and iteration cost exceeds token cost.
- Specialist routing by task type. The Fireworks data is instructive here — K3 outperformed on security and long terminal loops while Fable 5 won multilingual and data visualisation. Those are routable distinctions, not noise.
The teams getting the most out of 2026’s model landscape aren’t the ones who picked correctly. They’re the ones who built the switch.
Risks, Caveats and Red Flags
The distillation allegation. Michael Kratsios, director of the White House Office of Science and Technology Policy, has publicly accused Moonshot of building Kimi K3 by distilling Anthropic’s Claude Fable model. Moonshot has not conceded this. It’s an unresolved dispute with potential legal and procurement implications, and if you’re in a context where model provenance matters — government contracting, for instance — it belongs in your risk register.
Data routing and jurisdiction. All three companies operate primarily under Chinese jurisdiction. API calls traverse their infrastructure. Self-hosting resolves this; API usage does not. This is a compliance question, not a political one, and it has a clear technical answer if your requirements demand it.
Benchmark verification is genuinely thin. Qwen3.8-Max’s benchmark table is entirely vendor-reported with no independent replication published. GLM-5.2’s numbers arrived after launch from mixed sources. Only K3 has substantial third-party evaluation, and those evaluations disagree with each other.
Licensing uncertainty on Qwen. Precedent is not commitment. Until a Hugging Face repository with an actual license file exists, any plan that depends on Qwen3.8-Max weights is a plan built on a press release.
Quota mechanics change. GLM’s off-peak 1× multiplier promotion expires in September 2026, after which quota consumption roughly doubles off-peak and triples at peak. If you’re sizing a subscription on current effective throughput, model the post-promotion numbers too.
Release cadence is brutal. GLM-5 shipped in February, 5.1 in April, 5.2 in June. Kimi went K2.5 in January, K2.6 in April, K2.7 Code in June, K3 in July. Anything you conclude today has a shelf life measured in weeks.
What’s Coming Next
Z.ai’s next flagship is the most-anticipated near-term release. A JPMorgan research note, carried by Reuters and CGTN, points to an August 2026 window. Community leaks describe a model above one trillion parameters with the 1M context window carried forward and open weights, focused on long-running coding agents. Even the name is contested — GLM-5.3, GLM-5.5 and GLM-6 all appear in current discussion, with some reporting suggesting Z.ai may skip 5.3 and 5.4 entirely. As of August 4, Z.ai’s documentation still lists GLM-5.2 as the latest release. Vision support is the loudest community request by a wide margin.
Qwen3.8-Max weights are the nearest-term concrete event, promised for the week of August 10 alongside Qwen3.8-27B. The 27B checkpoint is the one to watch for practical on-premise deployment.
Kimi K3’s full technical report covering design, training methodology and benchmark reproduction was expected to follow the weight release.
I’ll update this comparison as each of those lands.
Frequently Asked Questions
Which is the best open-weight AI model right now?
On raw capability, Kimi K3 — it ranks fourth of 189 models on the Artificial Analysis Intelligence Index and first on Frontend Code Arena. On value for coding work, GLM-5.2, at less than a third of K3’s output cost with genuinely unrestricted MIT licensing. “Best” depends entirely on whether you’re optimising for capability or cost per resolved task.
Is Qwen3.8-Max actually open source?
Not yet. As of August 5, 2026 it is API-only. Alibaba committed to releasing weights the week of August 10 on Hugging Face and ModelScope, but no repository or license has appeared. Headlines calling it open source are describing a commitment, not a download.
Can GLM-5.2 process images?
No. GLM-5.2 has no visual encoder — it accepts text and code only. This was a deliberate architectural decision, not an oversight, and vision is the top community request for the next release.
How much does Kimi K3 cost?
$3 per million input tokens and $15 per million output tokens, with cached input at $0.30. That makes it the most expensive of these three by a wide margin, though still well below Claude Opus 5 at $5/$25 and Claude Fable 5 at $10/$50.
Can I run these models on my own hardware?
Realistically, not the flagships. Kimi K3 is a 1.56 TB download requiring 64+ accelerators. Qwen3.8-Max at 2.4T is a multi-node datacentre deployment. GLM-5.2 at ~753B total and ~40B active is the most tractable but still needs serious GPU infrastructure. The forthcoming Qwen3.8-27B is the realistic on-premise option.
Which model is best for coding agents?
Kimi K3 leads on Terminal-Bench 2.1 (88.3), Program Bench and SWE Marathon. GLM-5.2 offers the best cost-to-capability ratio at 81.0 on Terminal-Bench and 62.1 on SWE-bench Pro. Qwen3.8-Max sits between them at 86.6 and 67.7 while adding multimodal input. All three work with Claude Code, Cline and similar harnesses.
Do these models beat Claude and GPT?
On specific benchmarks, sometimes. Kimi K3 leads Frontend Code Arena over Claude Fable 5 and scores above Claude Opus 4.8 on the Artificial Analysis index. But Moonshot itself acknowledges K3 still trails Fable 5 and GPT-5.6 Sol overall, and independent testing shows mixed results by task category rather than a clean winner.
Which has the best license for commercial use?
GLM-5.2, unambiguously. Unrestricted MIT with no revenue thresholds or attribution requirements. Kimi K3’s custom license permits broad commercial use but adds conditions for companies above $20M revenue reselling it as a service. Qwen3.8-Max has no announced license at all.
Is it safe to use Chinese AI models for business data?
API calls route through infrastructure under Chinese jurisdiction. For personal projects and most startups this is a non-issue. For regulated industries it requires either legal review or self-hosting — which is exactly why GLM-5.2’s MIT license matters more than its benchmark scores for some organisations.
What’s the difference between total and active parameters?
Mixture-of-experts models activate only a fraction of their parameters per token. Kimi K3 has 2.8T total but activates roughly 50B (16 of 896 experts). Qwen3.8-Max has 2.4T total with ~95B active. GLM-5.2 has ~753B with ~40B active. Active parameters drive inference cost — which is why Qwen’s flagship is more expensive to serve than the larger K3.
Final Verdict
The most important thing this comparison reveals isn’t which model wins. It’s that the question changed.
Two years ago, the interesting question about open-weight models was how far behind they were. That question is now boring, because the answer is “close enough that it depends on your task.” Kimi K3 beat Claude Fable 5 on a blind human-preference leaderboard for front-end code. GLM-5.2 beat GPT-5.5 on real bug fixes at one-sixth the cost. These aren’t consolation-bracket results.
The interesting question now is what you’re actually buying. And the answer is different for each of these three in ways that have nothing to do with benchmark position:
- GLM-5.2 sells you independence. MIT weights, no conditions, no negotiation, the lowest price in the comparison. You’re buying the right to stop depending on a vendor.
- Kimi K3 sells you capability. The strongest scores here, at a price that reflects it, with license conditions that only bite if you get large.
- Qwen3.8-Max sells you breadth. Text, image and video in one model, strong agentic numbers, backed by Alibaba’s $53 billion infrastructure commitment — and, for now, a promise about weights rather than weights.
If I had to compress it to one line each: pick GLM-5.2 if you’re paying the bill yourself, Kimi K3 if capability is worth the premium, and Qwen3.8-Max if your agent needs eyes.
And whichever you pick, build the router. In a landscape where three frontier-adjacent models shipped in eight weeks, the ability to switch is worth more than the ability to choose correctly.
Last updated: August 5, 2026. Pricing, benchmarks and licensing in this space change weekly — verify current figures against official documentation before making procurement decisions. Sources include VentureBeat, MarkTechPost, Fortune, SCMP, Quartz, Artificial Analysis, Northflank and vendor documentation from Moonshot AI, Alibaba Qwen and Z.ai.
English 
Deutsch 
Español 
Italiano 
Français