Kimi K3 vs DeepSeek V4 Flash vs Qwen3.8-Max: Benchmarks, Price & Verdict

Short version: choose Kimi K3 for the strongest broad independent evidence, DeepSeek-V4-Flash-0731 for the lowest API price, and Qwen3.8-Max for the best current mix of multimodal reach and mid-tier cost.

All three models offer roughly one million tokens of context, reasoning controls and agent integrations. They are not interchangeable. Kimi K3 is a 2.8-trillion-parameter open-weight model with 104 billion active parameters and a $3/$15 API. DeepSeek V4 Flash uses a much smaller 284B/13B design and costs $0.14/$0.28. Qwen3.8-Max sits between them economically at $2/$6 while accepting text, images and video.

The evidence-led verdict: Kimi K3 leads the broad Artificial Analysis snapshot at 57 and the cleanest shared vendor-reported agent rows. DeepSeek V4 Flash scores 50 independently and is dramatically cheaper, but it is text-only and its current 0731 model is an API public beta rather than the model in DeepSeek’s consumer app. Qwen3.8-Max has a strong preliminary Arena Code/WebDev result and broad vendor-reported multimodal performance, but a comparable Artificial Analysis evaluation is still pending.

Qwen’s hosted production model is available now through QwenCloud and Qwen Chat. Alibaba says open weights will follow the week after launch. Until the checkpoint and license arrive, Qwen3.8-Max should be treated as a proprietary hosted model with an announced open-weight roadmap.

Updated August 3, 2026, Pacific time. This is a source-audited comparison, not a hands-on Kingy benchmark. Vendor scores, independent scores, current API models and downloadable checkpoints are labeled separately.

Qwen3.8-Max at a glance

Sources: the Qwen3.8-Max launch article, the QwenCloud model page and the official announcement.

What “2.4 trillion parameters, 95 billion active” actually means

Qwen3.8-Max is much larger on paper than most downloadable models, but it does not use all 2.4 trillion parameters for every token. A sparse mixture-of-experts model contains many specialist feed-forward networks and routes each token through only a subset. Alibaba says the active path is 95 billion parameters per token—about 4% of the total pool.

This design can deliver more learned capacity without multiplying the computation for every token by the full parameter count. It does not make Qwen3.8-Max a 95B model for deployment. A serving system generally needs fast access to the full expert pool, plus attention state, routing machinery, multimodal components, runtime buffers and redundancy.

These are derived lower bounds, not Alibaba hardware requirements. Qwen has not disclosed the training fleet, serving topology, native precision, quantization quality, tokens per second or time to first token. Even if the weight release arrives exactly as promised, practical self-hosting will be a multi-node project for hyperscalers, sovereign clouds, large enterprises and research consortia—not a desktop download. The separately promised Qwen3.8-27B may be far more consequential for local developers.

How good is Qwen3.8-Max?

Alibaba’s official table paints a model with unusual breadth rather than an undisputed lead everywhere. On coding and general agents, Qwen3.8-Max claims 86.6 on Terminal-Bench 2.1, 67.7 on SWE-bench Pro, 73.5 on FrontierSWE, 93.0 on PaperBench, 74.8 on CoWorkBench and 81.9 on WideSearch. On general reasoning it reports 92.6 on GPQA Diamond, 82.8 on IFBench, 60.2 on HealthBench and 92.9 on the 256K version of MRCR v2.

The strongest pattern is multimodal. Qwen reports 82.3 on MMMU-Pro, 95.2 without a code interpreter and 97.7 with one on MathVision, 91.9 on LogicVista, 91.5 on Parametric CAD Bench, 92.1 on OmniDocBench 1.5, 79.6 on CC-OCR-Bench v2 and 88.7 on VideoMMMU. It also claims 86.1 on OSWorld-Verified, a computer-use benchmark.

Those are not isolated wins on one kind of task. The model appears especially well positioned for workflows that combine documents, charts, screenshots, video, code and tool use. That is strategically more important than topping another short-answer knowledge test: enterprise work rarely arrives as clean text prompts.

Where Qwen leads in Alibaba’s table

• Research reproduction: 93.0 on PaperBench, ahead of the listed proprietary models.
• Instruction following: 82.8 on IFBench, above the comparison set.
• Professional domains: 60.2 on HealthBench, 73.2 on PLawBench and 58.3 on PRBench-Finance.
• Visual reasoning and perception: category-leading claims on LogicVista, HiPhO, several OCR/document tests, Parametric CAD Bench and multiple dense-perception evaluations.
• Long-context retrieval: 92.9 on MRCR v2 at 256K, close to GPT-5.6 Sol’s 93.8.

Where the frontier still leads

Alibaba’s own numbers show real gaps. Qwen’s 67.7 on SWE-bench Pro trails Fable 5’s 80.0. Its 56.6 on DeepSWE 1.1 is far below GPT-5.6 Sol at 73.0 and Fable 5 at 70.0. Qwen reaches 73.5 on FrontierSWE versus Fable 5 at 88.8, and 43.6 on Humanity’s Last Exam without tools versus Fable 5 at 53.3. On WebArena-Verified, Qwen reports 66.8 against Fable 5’s 71.3. Alibaba separately cautions that some Fable 5 results may involve fallback models, adding another reason to treat the exact gaps as directional.

That matters because the hardest repository-scale software tasks and long-horizon browser tasks often expose failures that easier coding benchmarks miss: incomplete environmental understanding, brittle recovery, poor prioritization and compounding mistakes. Qwen is close enough to be economically disruptive, but “close” is not the same as interchangeable on high-cost work.

The benchmark caveat is not a footnote

Launch tables are useful evidence, but this one is not a controlled tournament. Alibaba says some comparison numbers are best published results from different harnesses and warns that Fable 5 results can involve fallbacks. Qwen’s own evaluations variously use Claude Code, OpenCode, Qwen-Agent and other scaffolds; GPT models may use Codex. Several tests are internal Qwen benchmarks. Some results use code interpreters, web search, memory plugins or model judges. Alibaba also corrected selected task data and reran baselines in parts of the suite.

Even apparently identical labels can hide different timeouts, context limits, prompts, retry counts, sampling settings and tool permissions. Terminal-Bench 2.1, for example, used a five-hour timeout and average-over-ten procedure for Qwen, while comparison scores came from the best published results available. SWE-bench Pro used a Qwen-refined task set. WideSearch used different agents for Qwen and external models.

This does not make the scores worthless. It means the right conclusion is probabilistic: Alibaba has shown enough breadth to justify calling Qwen3.8-Max frontier class, but not enough standardized independent evidence to crown it the best model. As of August 3, Artificial Analysis had not published a production Qwen3.8-Max result. Arena had published a preliminary human-preference frontend score, which is useful but much narrower than a broad intelligence, latency and cost evaluation.

What the independent evidence says

The broadest comparable third-party data currently favors Kimi K3. Artificial Analysis scores Kimi K3 at 57 and DeepSeek V4 Flash 0731 at 50 on its Intelligence Index v4.1. Qwen3.8-Max has not yet completed that broad evaluation. It does have one useful independent signal: Arena’s human-preference Code/WebDev leaderboard lists Qwen at a preliminary 1,668, statistically overlapping Kimi K3 at 1,676 and ahead of DeepSeek V4 Flash High at a preliminary 1,577.

Sources: the independent Kimi K3 and DeepSeek V4 Flash 0731 result pages, plus the Arena Code/WebDev leaderboard, checked August 3, 2026.

The cost gap is striking but needs context. DeepSeek’s full Artificial Analysis run cost about 34 times less than Kimi’s even though DeepSeek generated far more output tokens. That shows how aggressively low token prices can absorb inefficient reasoning. It does not prove DeepSeek is cheaper per successful production task, where retries, human correction and failure cost matter.

Kimi’s 34.9-token-per-second output rate and 2.78-second time to first token are provider-specific measurements from its official API. Comparable Qwen and DeepSeek 0731 throughput figures were not available in the same methodology at publication, so this article does not invent a speed ranking.

Qwen3.8-Max vs Kimi K3 vs DeepSeek V4 Flash

Kimi K3 is the closest open-weight capability rival in scale and independently measured performance. DeepSeek-V4-Flash-0731 is far smaller and cheaper, but it is no longer merely an efficiency footnote: its official agent results improved over the April Preview, and Artificial Analysis scores the July API at 50. Treating all three as interchangeable “open models” misses major differences in modalities, licenses, deployment size, price and evidence.

Sources: Kimi K3’s official repository, Kimi’s model card and weights, DeepSeek’s official API changelog and Preview model card, plus the independent Kimi K3 and DeepSeek V4 Flash result pages.

The cleanest shared agent comparison

The following rows use matching benchmark names from each vendor’s release material. They are still not controlled reruns: agent harnesses, prompts, effort, timeouts and model versions may differ. Read the rank order as directional.

On this subset, Kimi K3 has the strongest coding-agent case. Its 57 Artificial Analysis score supplies the broader independent evidence Qwen still lacks; Qwen’s preliminary Arena result covers a narrower coding and web-development slice. Kimi’s drawbacks are price and license. At $3/M input and $15/M output, it costs 1.5 times as much as Qwen on input and 2.5 times as much on output. Its custom license requires a separate agreement for qualifying model-as-a-service businesses and imposes branding conditions on some very large products. “Open weight” does not necessarily mean unrestricted.

DeepSeek V4 Flash is about 8.45 times smaller than Qwen by total parameters and 7.3 times smaller on the active path. Qwen leads the shared vendor-reported agent rows, generally by two to four points, while Flash costs about 14 times less on uncached input and 21 times less on output. Artificial Analysis scores the 0731 API at 50, so it belongs in the capability conversation as well as the efficiency one. Qwen’s preliminary Arena result adds useful independent coding evidence, but the broader ranking remains open until it receives a comparable multi-domain evaluation.

The other large open-weight peers that matter

Qwen, Kimi and DeepSeek are the headline comparison, not the whole market. GLM-5.2 is the most important omitted peer: it is independently scored, MIT licensed, cheaper than Qwen and large enough to compete on serious agent workloads. MiniMax M3 adds another one-million-token multimodal option, while Mistral Large 3 remains a permissively licensed heavyweight.

Sources: the official GLM-5.2, MiniMax M3 and Mistral Large 3 model cards, plus Z.ai pricing and the independent GLM-5.2 and MiniMax M3 evaluations. MiniMax’s license and promotional pricing carry additional conditions; they are not equivalent to MIT or Apache 2.0.

Qwen’s price is the strategic weapon

Qwen3.8-Max launches at $2 per million input tokens, $6 per million output tokens and $0.25 per million implicit cached input tokens. Explicit cache creation is $2.50/M and explicit cache reads are $0.17/M. Those numbers are aggressive for a model Alibaba positions against the proprietary frontier.

Consider a substantial agent run consuming one million uncached input tokens and producing 200,000 output tokens:

The calculation uses published list prices and excludes caching, batch discounts, tool charges and provider-specific reasoning accounting. Closed-model prices come from the official Claude Fable 5 page, Claude API pricing table and GPT-5.6 Sol model page. Kimi, DeepSeek and GLM prices come from their respective Kimi, DeepSeek and Z.ai documentation.

Qwen is five times cheaper on input and more than eight times cheaper on output than Fable 5. It is 2.5 times cheaper on input and roughly 4.2 times cheaper on output than Opus 4.8. If independent tests place Qwen within striking distance on a company’s actual workload, the savings become too large to dismiss as a rounding error.

API price is not the same as self-hosting cost. Alibaba may be pricing strategically, operating at utilization levels outsiders cannot match or subsidizing adoption. A multi-node 2.4T deployment can be expensive even when only 95B parameters activate for each token. Companies should benchmark total throughput, latency, engineering labor, uptime and hardware utilization before assuming the downloadable model will be cheaper than the API.

Ease of use: the fastest path to each model

All three providers support familiar API formats, but the easiest route depends on whether the user wants a chat app, a coding agent, a multimodal API or downloadable weights.

A minimal OpenAI-compatible starting point

The same basic Python client can reach all three APIs. Choose one provider configuration, set the corresponding environment variable and keep the API key out of source code.

import os
from openai import OpenAI

PROVIDERS = {
    "qwen": {
        "key": "DASHSCOPE_API_KEY",
        "base_url": "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
        "model": "qwen3.8-max",
    },
    "kimi": {
        "key": "MOONSHOT_API_KEY",
        "base_url": "https://api.moonshot.ai/v1",
        "model": "kimi-k3",
    },
    "deepseek": {
        "key": "DEEPSEEK_API_KEY",
        "base_url": "https://api.deepseek.com",
        "model": "deepseek-v4-flash",
    },
}

provider = PROVIDERS["qwen"]  # change to "kimi" or "deepseek"
client = OpenAI(
    api_key=os.environ[provider["key"]],
    base_url=provider["base_url"],
)

response = client.chat.completions.create(
    model=provider["model"],
    messages=[{"role": "user", "content": "Compare two technical proposals."}],
)
print(response.choices[0].message.content)

This example proves connection, not production readiness. Add each provider’s documented reasoning controls, streaming, retries and tool schema for a real application. During a tool-calling loop, the safest cross-provider rule is to preserve the complete assistant message returned by the API. Qwen and Kimi also expose richer multimodal paths than the text-only snippet above.

Official setup sources: Qwen3.8-Max API and coding-assistant guide, the Kimi K3 quickstart, and the DeepSeek thinking-mode guide.

Which one is easiest?

Kimi offers the clearest no-code path because the same model appears across Kimi.com, Kimi Work and Kimi Code. Qwen is the easiest choice for a hosted multimodal API because QwenCloud documents image, video, tools, caching and three compatible protocols in one place. DeepSeek is the cheapest text API to start testing, but users who specifically need the 0731 upgrade must use the API rather than assume the consumer app is equivalent.

The autonomous-work demos: impressive, but not reliability statistics

Alibaba’s most provocative claims are not benchmark scores. They are long-running projects intended to show that Qwen3.8-Max can maintain state, recover from errors and improve work over days.

In the public oh-my-cli repository, Qwen says the model worked for roughly 16 days and produced 265 commits, 127 pull requests and 151 issues by July 30. In a research-reproduction project, Alibaba reports about 125 continuous hours, 7,600 lines of code, more than 1,100 actions and 33 GPU-training rounds. The model first reproduced six findings from a paper, then explored 18 proposed improvements and raised an AIME24 result from 49.58% to 52.29%.

Other launch cases include roughly 500 turns of chip optimization, a 365-day simulated e-commerce operation and a quant-research workflow with hundreds of subagents and thousands of backtests. The breadth is notable: Alibaba is selling a coworker that can span code, browsers, graphics, documents and domain tools, not merely a chatbot with a larger prompt window.

These remain curated vendor demonstrations. They do not reveal total token use, compute cost, human intervention, rejected runs, error rate, sandbox design or whether an independent team can reproduce the result. Counts of commits and tool calls are measures of activity, not necessarily value. The right takeaway is that Qwen has demonstrated credible long-horizon possibility; it has not yet published a reliability curve for unsupervised multi-day work.

What Qwen3.8-Max means for closed frontier labs

1. Benchmark intelligence alone cannot sustain a premium

If Qwen’s independent results land near Alibaba’s claims, closed labs will have to justify four-to-eightfold output-price premiums with the rest of the product: lower error rates, predictable latency, stronger security controls, enterprise governance, indemnification, support, auditability, better agents and superior user experience. A small average benchmark advantage can still be worth a great deal on high-stakes work, but customers will demand evidence that the advantage survives their own tasks.

2. The moat moves from the checkpoint to the system

Downloadable frontier weights reduce switching costs. They let hyperscalers, governments and enterprises control deployment, fine-tune privately, audit behavior, build specialized guardrails and avoid a single provider. Closed labs still possess meaningful advantages in training infrastructure, post-training recipes, inference optimization, proprietary data, developer tools and distribution. The model becomes one layer of the moat rather than the entire moat.

3. Model routing becomes the default architecture

The market now offers several quality-cost points rather than one neat ladder. DeepSeek V4 Flash pairs a strong independent score with exceptionally low prices. GLM-5.2 sits between Flash and Qwen on price and independent performance. Qwen3.8-Max adds broader multimodal and vendor-reported agent strength; Kimi K3 provides a separately validated large open-weight option. Premium closed models remain attractive for difficult repository work, autonomous cyber evaluation, maximum reliability and integrated enterprise support. The rational buyer routes each task to the cheapest model that clears a measured quality threshold.

4. Qwen’s ecosystem may matter more than one leaderboard position

The ATOM Report estimated that Qwen-family models had accumulated roughly 942 million Hugging Face downloads and supplied 69% of newly created model adaptations in February 2026. Downloads are noisy and do not equal active users, but the derivative ecosystem is already enormous. A strong base checkpoint can spread through fine-tunes, quantizations, local tools and national infrastructure faster than a hosted model whose behavior remains controlled by one vendor.

5. “Closed versus open” is no longer a capability ranking

Stanford’s 2026 AI Index measured a narrowing gap between the best closed and open models and a compressed U.S.–China frontier. Qwen3.8-Max strengthens that trend, but it does not prove closed labs are obsolete. Alibaba’s own table shows Fable 5 and GPT-5.6 Sol retaining material leads on several hard software-engineering and reasoning tests. The defensible conclusion is narrower and more important: licensing model is becoming a poor proxy for intelligence.

Open weight is not the same as open source—or full transparency

Alibaba describes this as its first open Max model, but the precise deliverable matters. Publishing weights allows inference, modification and analysis. It does not automatically publish the training data, data licenses, training code, full post-training recipe or compute history. Weights alone therefore do not settle either open-source status or reproducibility; the license, accompanying code, documentation and data disclosures determine which freedoms and evidence users actually receive.

Qwen3.8-Max also has no public license at launch. Older Qwen releases have used permissive terms, but it would be irresponsible to assume the same license before the actual repository appears. Kimi K3 shows why: its weights are downloadable, yet its custom license imposes conditions on certain high-revenue services and large products. DeepSeek V4 Flash is simpler because its published checkpoint is MIT licensed.

The promised release will only be complete enough to evaluate when Alibaba publishes the model card, exact license, configuration, tokenizer, precision formats, inference instructions and safety documentation. Until then, “open weight next week” is a roadmap commitment.

The unresolved safety and governance questions

A 95B-active multimodal coding model with strong computer use deserves more than capability tables. Alibaba has not published a dedicated Qwen3.8-Max system card describing cyber, biological, autonomy, privacy or deception evaluations. The separate Qwen3Guard project is useful infrastructure, but it is not evidence that Qwen3.8-Max has equivalent or tamper-resistant safeguards.

The International AI Safety Report 2026 captures the open-weight trade-off: downloadable models improve competition, auditing and defensive research, while also making safeguards easier to remove and releases effectively irreversible. The UK AI Security Institute’s preliminary Kimi K3 cyber evaluation offers a useful precedent. Kimi showed meaningful offensive capability but remained behind leading closed models on difficult autonomous tasks, and its safeguards did not reliably stop offensive attempts.

Before treating Qwen3.8-Max as infrastructure, organizations should ask whether it has been independently evaluated for dangerous capabilities, what guardrails survive checkpoint modification, which data-governance commitments apply, and how the model behaves under long-running tool access. Openness can enable better scrutiny, but it transfers more safety engineering to the deployer.

Who should use which model?

• Choose Qwen3.8-Max when the workload combines images, video, documents, code and tools; one-million-token context matters; and premium closed-model pricing is difficult to justify. Run a private task suite before migration.
• Choose Kimi K3 when independently measured frontier performance and downloadable weights matter more than API price, and its commercial license fits the deployment.
• Choose DeepSeek V4 Flash when the workload is text-only and price or a 13B active path dominates. The April Preview weights are MIT licensed, but the improved 0731 model was API-only at this update; do not treat them as the same artifact. Its independent score is competitive, but test the actual task rather than assuming the lower price decides the ranking.
• Consider GLM-5.2 when a one-million-token, MIT-licensed model with a 51 Artificial Analysis score and prices below Qwen fits the workload; it lacks Qwen’s native image and video breadth.
• Keep a premium closed model in the route for tasks where a modest success-rate difference produces large downstream costs, or where enterprise support, liability, governance and the provider’s full tool ecosystem matter.
• Wait for Qwen3.8-27B if the real objective is broad local deployment rather than hyperscale self-hosting.

The practical evaluation should measure task completion, human correction time, catastrophic error rate, latency, total tokens, tool calls and cost—not one benchmark average. For agents, a 3% higher pass rate can be enormously valuable on expensive work; on routine extraction, it can be a waste of money.

Final verdict

There is no universal winner here; there are three different kinds of value. Kimi K3 has the strongest broad independent case in this comparison: an Artificial Analysis Intelligence Index score of 57, solid shared agent results and a downloadable checkpoint. The trade-offs are its $3/$15 API price, 34.9-token-per-second measured output speed and a custom license that needs closer review than MIT.

DeepSeek-V4-Flash-0731 is the value winner. Its $0.14/$0.28 regular API price is dramatically lower than the other two, while its Artificial Analysis score of 50 remains competitive. It also has the clearest permissive path through the MIT-licensed Preview checkpoint. The catches matter: the current 0731 API is text-only, its public checkpoint and consumer app are not the same model version, and reasoning runs can be verbose.

Qwen3.8-Max is the most balanced hosted option for multimodal work. It accepts text, images and video, costs $2/$6 per million tokens and posted a preliminary 1668 ± 18 score in Arena’s Code/WebDev leaderboard—close to Kimi K3’s 1676 ± 12 on that narrow test. It still lacks a broad Artificial Analysis result, and Alibaba’s promised weights, license and practical self-hosting requirements were not yet verifiable when this update was published.

Choose Kimi K3 when independent breadth matters most, DeepSeek when cost or an MIT path dominates, and Qwen when one hosted model must combine coding with visual and video inputs. For production routing, run all three against the same private task set, tool harness, reasoning budget and latency target. Their headline benchmark numbers were not produced under one controlled protocol.

Kimi K3 vs DeepSeek V4 Flash vs Qwen3.8-Max FAQ

Which model is cheapest?

DeepSeek-V4-Flash-0731 is the clear API price leader at $0.14 per million uncached input tokens and $0.28 per million output tokens at current regular rates. Qwen3.8-Max costs $2/$6, while Kimi K3 costs $3/$15. Cached-input rates, output length and retry rates can change the total cost of a completed task, so token prices are only the starting point.

Which model is easiest to use?

For no-code multimodal work, Qwen3.8-Max has the simplest all-in-one path through Qwen Chat. For API migration, all three offer OpenAI-compatible endpoints; DeepSeek is usually the easiest budget experiment, but the 0731 model is API-only and its thinking-mode tool loops require reasoning history to be preserved. Kimi’s web app is polished, while self-hosting either Kimi or DeepSeek requires data-centre-class hardware.

Is Qwen3.8-Max open source?

No—not yet. It is available as a hosted model, and Alibaba said open weights would arrive the following week. The checkpoint and license were not public when this update was prepared. Downloadable weights alone would not establish whether the release meets every definition of open source.

How large are the three models?

Qwen3.8-Max has 2.4 trillion total parameters and activates 95 billion per token. Kimi K3 is larger at 2.8T total and 104B active. DeepSeek V4 Flash is much smaller at 284B total and 13B active. Parameter counts do not translate directly into quality, speed or hardware cost because architectures, precision and serving stacks differ.

Can these models run locally?

Kimi K3 and the DeepSeek V4 Flash Preview have downloadable weights, but both are data-centre-scale deployments rather than ordinary workstation models. Qwen3.8-Max weights had only been promised, not released, at publication time. Most readers should start with hosted APIs before considering multi-accelerator infrastructure.

Is Qwen3.8-Max better than Kimi K3?

It depends on the workload. Qwen is cheaper and natively accepts image and video inputs. Kimi has the stronger broad independent record, scoring 57 on Artificial Analysis. In Arena’s preliminary Code/WebDev leaderboard, their scores were close—1668 ± 18 for Qwen and 1676 ± 12 for Kimi—but that frontend test is narrower than a full intelligence evaluation.

Is Qwen3.8-Max better than DeepSeek V4 Flash?

Qwen is the stronger fit for multimodal tasks and leads several vendor-reported agent results. DeepSeek is the price and efficiency choice: it is text-only, activates 13B parameters, costs far less per token and scores 50 on Artificial Analysis. Do not mix the July 31 0731 API, the unchanged consumer app and the downloadable Preview checkpoint as if they were identical.

Which one should a developer choose?

Start with DeepSeek when budget dominates, Kimi when broad independent performance matters most, and Qwen when visual or video inputs are part of the workflow. Then benchmark the finalists on your own tasks with the same prompts, tools, reasoning settings and latency limits.

Primary and independent sources

• Qwen3.8-Max official launch article
• QwenCloud specifications, features and pricing
• Alibaba Qwen launch announcement
• Qwen’s public oh-my-cli autonomous-development trace
• Kimi K3 official repository and technical material
• Kimi K3 model card, weights and license
• DeepSeek V4 Flash Preview model card and MIT-licensed weights
• DeepSeek API pricing
• GLM-5.2 official model card and weights
• MiniMax M3 official model card and weights
• Mistral Large 3 official model card and weights
• Artificial Analysis: Kimi K3
• Artificial Analysis: DeepSeek V4 Flash
• Artificial Analysis: GLM-5.2
• UK AI Security Institute: preliminary Kimi K3 cyber assessment
• International AI Safety Report 2026
• Stanford AI Index 2026: technical performance
• ATOM Report on the global open-model ecosystem