Wealth Engine

Tag: Open Source AI

  • Tencent Hy4: 770B Open Weights, 82% Cheaper Output Than Kimi K3

    Tencent Hy4 preview is a 770-billion-parameter open-weight model released under Apache 2.0 on August 28, 2026, with 49B active parameters and a 1,048,576-token context window. Tencent Cloud prices it at $0.834 per million input tokens and $2.501 per million output — roughly one-eighth the output cost of GPT-5.6 Sol and 82% below Kimi K3. Weights are on Hugging Face.

    Tencent open-sourced its largest model to date on Friday, and the interesting number is not the parameter count. It is the price tag attached to weights anyone can download, modify and resell.

    That combination — frontier-adjacent scores, permissive licensing, and output tokens at $2.501 per million — is the part that moves money.

    What is Tencent Hy4 preview?

    Tencent Hy4 preview is a mixture-of-experts language model with 770B total parameters, of which 49B activate per token. Tencent released it on August 28, 2026, published the weights on Hugging Face under the Apache 2.0 license, and shipped it simultaneously into its own consumer and developer products.

    According to Tencent’s announcement, the model is live in WorkBuddy, CodeBuddy, Yuanbao and ima, with API access through Tencent Cloud TokenHub and OpenRouter.

    WorkBuddy and CodeBuddy are free for two weeks from launch. That is a customer-acquisition subsidy, not a pricing model.

    The architecture, briefly

    The Hugging Face model card lists 78 layers — the first a dense FFN, the remaining 77 MoE — with 256 routed experts plus one shared expert per MoE layer, and top-8 routing per token.

    Vocabulary size is 120,832. Attention is what Tencent calls “Gated Sparse Attention with IndexCache,” reusing sparse indices across layers to keep the million-token window affordable to serve.

    There is also a built-in multi-token-prediction layer of 10B parameters (0.7B activated) for speculative decoding. Tencent says system-level optimization lifted end-to-end throughput 31.8% against its own baseline.

    How much does Tencent Hy4 cost?

    Tencent Cloud lists $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million on cache hits. In renminbi terms Tencent quotes ¥6 input and ¥18 output per million tokens.

    Those are the headline numbers, and they are aggressive against every comparable model.

    Per KuCoin’s summary of the launch materials, that input price is 25% below GLM-5.3 and 70% below Kimi K3; on output it is 36% below GLM-5.3 and 82% below Kimi K3.

    Model Input / 1M Output / 1M Context Weights
    Tencent Hy4 preview $0.834 $2.501 1,048,576 Apache 2.0
    GPT-5.6 Sol (base tier) $4.00 $20.00 n/d Closed
    GPT-5.6 Sol (long context) $8.00 $30.00 n/d Closed

    On output — the expensive half of any agentic workload, where the model writes code, calls tools and re-reads its own work — Hy4 runs at roughly one-eighth of GPT-5.6 Sol’s base rate.

    For a coding agent burning 50 million output tokens a month, that is the difference between about $1,000 and about $125. The comparison only holds if the cheaper model finishes the job in a similar number of tokens, which is exactly the assumption worth testing.

    How does Hy4 score on benchmarks?

    Hy4 posts strong software-engineering numbers and slightly trails the closed frontier on reasoning. Tencent reports 92.3 on GPQA Diamond, 65.7 on SWE-Bench Pro, 82.9 on SWE-Bench Multilingual, 62.9 on SkillsBench v1.1, and 64.3 on Deep-SWE.

    The Deep-SWE result is the standout: 64.3 against 28.0 for the previous generation. That is not incremental.

    Against closed models the gap is real but narrow. Hy4 scores 92.3 on GPQA Diamond versus GPT-5.6 Sol’s 94.6, and 85.4 on Terminal-Bench versus Sol’s 88.8.

    • GPQA Diamond: Hy4 92.3 — GPT-5.6 Sol 94.6
    • Terminal-Bench: Hy4 85.4 — GPT-5.6 Sol 88.8
    • SWE-Bench Multilingual: Hy4 82.9
    • SWE-Bench Pro: Hy4 65.7
    • Deep-SWE: Hy4 64.3, up from 28.0

    Every one of those figures is vendor-reported. None has been independently reproduced at the time of writing, and the model has been public for about a day.

    Is Hy4 better than GLM-5.3 and Kimi K3?

    Marginally, on Tencent’s own evidence. Tencent ran a blind evaluation with 163 internal experts across 203 real engineering tasks. Hy4 preview averaged 2.99 out of 4.00, against 2.94 for Kimi K3 and 2.92 for GLM-5.3.

    A 0.05-point spread on a four-point scale is not a capability gap. It is a tie with a favorable rounding.

    The win-rate breakdown is more honest about how close this is. Against GLM-5.3, Hy4 won 46.8% of comparisons, drew 12.8% and lost 40.4%. Against Kimi K3: 51.2% wins, 7.9% draws, 40.9% losses.

    So Hy4 loses roughly two of every five head-to-head comparisons against models that were already open. And per KuCoin’s read of the same materials, Hy4 “did not lead comprehensively in public benchmarks” and lags GLM-5.3 in code and cybersecurity tests.

    The differentiator here is price and license, not raw capability. That is worth saying plainly, because the launch framing does not. If you are choosing between cheap open coders, our GLM-5.3 vs DeepSeek V4 Pro comparison and the GLM-5.3-Flash vs Qwen3.8-Flash-Next breakdown cover the alternatives.

    Should you self-host Hy4 or use the API?

    Self-hosting is viable at a scale that would have required a cluster a year ago. Tencent documents deployment on a single eight-GPU node using the FP8 quantized variant, with prebuilt containers for vLLM and SGLang at tensor-parallel size 8.

    The full BF16 checkpoint and an FP8 build are both published, along with the AngelSlim toolkit for further compression.

    The API case is stronger for anyone under roughly 100 million tokens a month. One eight-GPU node of current-generation accelerators plus the engineer who babysits it will not come in under $2.501 per million output tokens at low volume.

    The self-host case is stronger for three groups: teams with data-residency constraints, teams already running GPU capacity at low utilization, and teams that want to distill Hy4 into something smaller. Apache 2.0 permits all three without a negotiation. Our earlier analysis of Qwen3.8-Max open weights versus API walks through that math in detail.

    Who wins and loses financially?

    The clearest loser is anyone selling mid-tier closed inference. If a 770B open model at $2.501 output holds up in production, the price umbrella over $20-per-million output tiers gets thinner.

    Kimi K3 is the most directly exposed. An 82% output-price gap against a model that wins only 51.2% of blind comparisons is a hard position to hold.

    Tencent wins on distribution, not on API margin. At ¥6 per million input, the API is a loss leader that routes developers toward Tencent Cloud, WorkBuddy and CodeBuddy — the same playbook that made cheap Chinese inference a strategic instrument rather than a business line. We covered the last turn of that cycle when DeepSeek raised prices by up to 1,100%.

    The GPU vendors win either way. Open weights that need eight accelerators per node create hardware demand that a closed API never surfaces on anyone else’s balance sheet.

    What are the catches?

    Three, and Tencent names one of them itself.

    The company acknowledges in its release notes that the model spends “longer than necessary reasoning” and over-verifies its own work. On a consumption-priced endpoint, verbosity is a bill. A model that is 82% cheaper per token but writes twice as many tokens is 64% cheaper, not 82%.

    Second, serving is thin. OpenRouter shows a single provider — Tencent Cloud — with 3.16-second P50 latency, 38 tokens per second, and 98.65% availability over three days. There is no failover route.

    Third, the context window is a spec, not a guarantee. The model accepts 1,048,576 input tokens but caps completions at 64,000, and nothing in the release claims uniform recall across the full window. For a like-for-like look at long-context pricing, see our cheapest 1M-context model comparison.

    And it is called “preview” for a reason.

    Frequently asked questions

    Is Tencent Hy4 preview free to use?

    The weights are free under Apache 2.0. The hosted API is not — it costs $0.834 per million input tokens and $2.501 per million output. WorkBuddy and CodeBuddy are free for two weeks from the August 28 launch.

    Can I use Hy4 commercially?

    Yes. Apache 2.0 permits commercial deployment, modification, distillation and redistribution without a separate license negotiation or revenue threshold.

    What hardware do I need to run Hy4?

    Tencent documents a single eight-GPU node using the FP8 quantized build, served through vLLM or SGLang at tensor-parallel size 8. Minimum memory figures are not published.

    How big is the context window?

    1,048,576 tokens of input, with completions capped at 64,000 tokens.

    Is Hy4 better than GPT-5.6 Sol?

    Not on published benchmarks. Hy4 scores 92.3 on GPQA Diamond against Sol’s 94.6, and 85.4 on Terminal-Bench against 88.8. It is cheaper by roughly 8x on output.

    Where can I download the weights?

    Hugging Face at tencent/Hy4-preview, with code and deployment instructions on GitHub.

    Have the benchmarks been independently verified?

    No. All published scores are vendor-reported as of August 29, 2026.

    The bottom line

    Hy4 preview is not a capability breakthrough. It wins its own blind evaluation by 0.05 points and loses 40% of head-to-head comparisons against models that were already open-weight.

    It is a pricing event. Tencent shipped near-parity performance under Apache 2.0 at 82% below Kimi K3’s output rate and roughly one-eighth of GPT-5.6 Sol’s, and put the weights on Hugging Face the same day.

    Tencent’s own README calls it “another step change in capability — the largest generation-over-generation gain we’ve measured.” The blind-evaluation table does not support that framing. The invoice does.

    For anyone running high-volume agentic workloads, Hy4 is worth a benchmark run this week — with token-consumption logging turned on, because the verbosity Tencent admits to is where the savings go to die. For anyone selling inference above $20 per million output tokens, the floor moved again.

    Sources

  • Nvidia Hugging Face Acquisition: $12.9 Billion at 80x Revenue

    The Nvidia Hugging Face acquisition values the open-source model hub at $12.9 billion, according to The Information — roughly 80 times its ~$150 million in annualized revenue. Hugging Face turned down a $500 million Nvidia investment at a $7 billion valuation less than a year ago. Neither company has confirmed the deal. It would be Nvidia’s second-largest acquisition ever, behind the $20 billion Groq purchase.

    How much is Nvidia paying for Hugging Face?

    Nvidia has agreed to pay approximately $12.9 billion for Hugging Face, The Information reported on August 27, citing a person familiar with the transaction. The deal is agreed but not signed. It can still collapse.

    CNBC and Fortune both matched the report the same day. Business Insider first reported Nvidia’s takeover interest.

    Neither Nvidia nor Hugging Face responded to requests for comment, per Quartz. That silence matters — nothing here is a signed, disclosed transaction yet.

    What Hugging Face was worth before

    Hugging Face last priced itself at $4.5 billion in a 2023 round led by Salesforce Ventures, with Alphabet’s GV, IBM Ventures and — notably — Nvidia participating, according to TechCrunch.

    In late 2025, Nvidia offered $500 million at a $7 billion valuation. Hugging Face said no. TechCrunch reports the company declined because it did not want a single dominant investor.

    Roughly nine months later, it is selling outright to that same investor for nearly double the valuation it rejected.

    Date Event Valuation
    2023 Series funding led by Salesforce Ventures $4.5 billion
    Late 2025 $500M Nvidia investment offer — declined $7 billion
    Aug 27, 2026 Reported acquisition agreement $12.9 billion

    Why is Nvidia buying an open-source model hub?

    Nvidia is buying distribution, not revenue. Hugging Face is where developers publish, discover and download open models — Tom’s Hardware calls it a “GitHub-like repository” for AI. Owning the shelf is worth more to Nvidia than the $150 million the shelf currently earns.

    The strategic timing is not subtle. Nvidia’s largest customers are building silicon that competes with its own.

    The lock-in play

    Hugging Face’s Inference Endpoints today support AWS Inferentia, AMD Instinct, Google TPU, Intel CPUs and Nvidia accelerators, per Tom’s Hardware. It is deliberately vendor-neutral.

    Under Nvidia, that neutrality is the first thing analysts expect to erode. If the default deployment path for every popular open model points at CUDA, Nvidia defends its installed base at the exact layer where switching decisions get made.

    That threat is real. OpenAI, Google, Amazon and Anthropic are all shipping or funding custom accelerators — see our coverage of OpenAI’s Jalapeño chip and the Broadcom debt package funding Anthropic’s silicon.

    The cloud re-entry play

    Nvidia scaled back DGX Cloud roughly a year ago, TechCrunch notes. Hugging Face gives it a consumer-facing compute surface again — and somewhere to route the capacity Nvidia has committed to but not fully sold.

    That is the least-discussed part of the rationale and possibly the most financially concrete one.

    Is $12.9 billion too much for $150 million in revenue?

    On the numbers, yes — by any conventional standard. Tom’s Hardware puts the deal at roughly 80 times forward revenue. Software acquisitions at 15–20x are already considered rich.

    Hugging Face’s revenue is growing fast. TechCrunch reports it moved from about $100 million to about $150 million annualized in roughly two months, and Tom’s Hardware says paying subscribers doubled in the first half of 2026. CEO Clem Delangue told TechCrunch last month the company was “close to profitability.”

    Here is the skeptical read. Nvidia is paying a strategic premium for neutrality it intends to end. The moment developers believe Hugging Face is a CUDA storefront rather than a Switzerland, some of them leave — and the asset Nvidia bought is worth less than the asset it paid for. AMD, Google and the open-weights community have every incentive to fund an alternative registry.

    Ten-year-old infrastructure businesses with $150 million in revenue do not usually command $12.9 billion. They command it when the buyer is defending a franchise.

    How does this fit Nvidia’s acquisition spree?

    It is the second-largest deal Nvidia has ever done, and the third multi-billion-dollar AI purchase in nine months. Nvidia has stopped behaving like a component supplier and started behaving like a platform consolidator.

    Target Reported price Announced What it buys
    Groq ~$20 billion Dec 2025 Inference architecture (LPU)
    Hugging Face $12.9 billion Aug 2026 Open-model distribution
    Poolside ~$6 billion Aug 2026 Model training capability

    The Groq deal — about $20 billion, reported by CNBC in December 2025 — was Nvidia’s largest on record. We covered the $6 billion Poolside purchase earlier this month.

    Nvidia can afford all of it in cash. Its Q2 fiscal 2027 results, for the quarter ended July 26, 2026, show $96.2 billion in revenue, up 106% year over year, and $59.7 billion in GAAP net income. Cash, marketable debt and marketable equity securities totaled roughly $99.3 billion.

    Why this matters

    Three things follow from this deal, and none of them are about Hugging Face.

    • The competitive threat is now priced. Nvidia is spending real money to defend against customers building their own chips. That is an admission the threat is material.
    • Open-source AI just got an owner. The default distribution point for open models moves inside a hardware vendor. Expect immediate pressure for a neutral alternative.
    • Strategic multiples are back. Eighty times revenue is a 2021-style number appearing in 2026, funded by operating cash rather than cheap debt.

    For investors, the read is about defensive capital allocation. Nvidia guided to about $108 billion for the current quarter — excluding any China data center compute revenue. A company growing that fast does not spend $12.9 billion on a $150 million business unless it sees a hole in the moat.

    It also fits a pattern of Nvidia paying to control adjacent chokepoints, much as Stripe paid over $7 billion for OpenRouter to sit on the AI token toll road. Note, too, that Nvidia recently cut its OpenAI data center guarantee from $250 billion to $120 billion — capital is being redirected, not simply added.

    This post is reporting and analysis, not financial advice.

    Frequently asked questions

    Is the Nvidia Hugging Face acquisition confirmed?

    No. The Information reported an agreement on August 27, 2026, and CNBC, Fortune and TechCrunch matched it. Neither company has commented publicly, and the deal is not signed.

    How much revenue does Hugging Face generate?

    Roughly $150 million annualized, up from about $100 million two months earlier, according to TechCrunch. The $12.9 billion price is about 80 times that figure.

    Why did Hugging Face reject Nvidia before?

    It declined a $500 million investment at a $7 billion valuation in late 2025 because it did not want a single dominant investor, TechCrunch reported.

    Will Hugging Face still support AMD and Google chips?

    Unknown. Its Inference Endpoints currently support AWS Inferentia, AMD Instinct, Google TPU, Intel CPUs and Nvidia accelerators. Nvidia has not said whether that continues.

    Is this Nvidia’s biggest acquisition?

    No. The roughly $20 billion Groq deal announced in December 2025 remains its largest, per CNBC. Hugging Face would rank second.

    Could regulators block the deal?

    No formal review has been reported. Antitrust scrutiny is plausible given Nvidia’s accelerator share and the platform’s role in model distribution, but nothing has been filed publicly.

    Can Nvidia pay cash?

    Comfortably. It reported $59.7 billion in GAAP net income in a single quarter and about $99.3 billion in cash and marketable securities as of July 26, 2026.

    The bottom line

    Nvidia is paying roughly 80 times revenue to own the front door of open-source AI. The financial case is thin; the defensive case is obvious.

    Watch three things next: whether the deal is actually signed, whether Hugging Face keeps supporting rival accelerators, and how quickly a neutral competitor gets funded. The first tells you if this is real. The second and third tell you whether $12.9 billion bought a moat or a melting asset.

    Sources

  • GLM-5.3 vs DeepSeek V4 Pro: Which Open-Weight Coder to Buy

    GLM-5.3 vs DeepSeek V4 Pro comes down to one thing: you can download DeepSeek today. Its MIT-licensed weights shipped August 12, 2026, with a 1M-token context, and DeepSeek’s own API lists $0.435 per million input tokens against $0.87 output. GLM-5.3 landed two days later with stronger cyber scores, no published per-token price, and weights held back roughly two weeks.

    Two of the most consequential open-weight coding models of the year shipped 48 hours apart. Both vendors published benchmark tables claiming frontier-class performance. Neither number has been independently replicated.

    Here is what the money actually says.

    What shipped in the GLM-5.3 vs DeepSeek V4 Pro week?

    DeepSeek V4 Pro 0813 went generally available on August 12, 2026, after a preview build dated April 24. Z.ai launched GLM-5.3 on August 14. Both target agentic coding and terminal work. Only one of them can be run on your own hardware right now.

    DeepSeek V4 Pro 0813: 1.6 trillion parameters, MIT license

    V4 Pro is a mixture-of-experts model with 1.6 trillion total parameters and 49 billion active per token, according to TechTimes’ launch coverage. Context window is 1,048,576 tokens. The license is MIT — the most permissive terms in the frontier tier.

    Artificial Analysis ranks it third out of 1,075 models evaluated on its Intelligence Index, with a score of 53, and measures output at 75.4 tokens per second with a 1.69-second time to first token.

    GLM-5.3: same base model, rebuilt post-training

    Z.ai did something unusual. GLM-5.3 reuses the 743-billion-parameter base model from GLM-5.2 with no new pre-training run, per MarkTechPost’s technical breakdown. Every gain came from post-training.

    The gains are not small. Terminal-Bench 3.0 went from 4.6 to 28.3. DeepSWE v1.1 climbed from 46.2 to 66.9. That is a roughly 45% relative jump on agentic software engineering from post-training alone.

    Which is better for coding, GLM-5.3 or DeepSeek V4 Pro?

    On raw self-reported coding scores, DeepSeek V4 Pro leads. It claims 62.7 on DeepSWE and 87.9 on Terminal-Bench 2.1. GLM-5.3 reports 66.9 on DeepSWE v1.1 and 28.3 on Terminal-Bench 3.0. The benchmarks are different versions, so the comparison is not apples to apples — and that is the whole problem.

    The benchmark numbers nobody has verified

    DeepSeek’s self-reported 87.9 on Terminal-Bench 2.1 would place it first in the world. It does not appear on the benchmark’s own board.

    The official Terminal-Bench 2.1 leaderboard is topped by Claude Code running Fable 5 at 83.8% ± 1.2%, followed by Codex with GPT-5.5 at 83.1% and Cursor CLI with Grok 4.5 at 79.3%. No DeepSeek entry exists.

    TechTimes noted the same gap: V4 Pro’s scores “have not been independently replicated by any third-party evaluator as of publication.”

    Independent evaluator Artificial Analysis runs its own Terminal-Bench v2.1 harness and gets different absolute numbers again: GPT-5.6 Sol at xhigh effort scores 89.5%, Claude Opus 5 at max effort 89.1%, Grok 4.6 at 88.4%.

    Three sources, three scales, zero cross-comparability. Treat every vendor coding number as a marketing claim until a third party runs the harness.

    Where GLM-5.3 actually wins

    Security. GLM-5.3 scores 84.5% on CyberGym, up from 77.2%, ahead of Mythos 5 at 83.8% and GPT-5.6 Sol at 83.6%. DeepSeek V4 Pro self-reports 83.3%.

    ExploitBench more than doubled, from 24.4% to 54.4%. Z.ai says the model surfaced 2,436 vulnerabilities across 269 open-source projects, 1,097 of them rated critical or high severity.

    If your workload is vulnerability triage or defensive security automation, that is the strongest open-weight number on the board.

    How much does each model cost per million tokens?

    DeepSeek is cheaper — but only if you buy direct. The same model costs roughly 2.8x more input and 2.8x more output through a marketplace. GLM-5.3 has no published per-token rate at all; Z.ai’s price table still lists GLM-5.2. That pricing opacity is a real cost.

    Metric DeepSeek V4 Pro 0813 GLM-5.3
    Launch date Aug 12, 2026 (GA) Aug 14, 2026
    Input / 1M (vendor direct) $0.435 cache miss Not published (GLM-5.2: $1.40)
    Cached input / 1M $0.003625 Not published (GLM-5.2: $0.26)
    Output / 1M (vendor direct) $0.87 Not published (GLM-5.2: $4.40)
    Input / output via OpenRouter $1.218 / $2.436 GLM-5.2: $0.50 / $3.15
    Parameters 1.6T total / 49B active 743B base (shared with 5.2)
    Context window 1,048,576 tokens Not disclosed
    License MIT, weights live Weights ~2 weeks post-launch
    Subscription option None $18 / $80 / $168 per month

    Sources: TechTimes, OpenRouter, OpenRouter GLM 5.2.

    The marketplace markup is the hidden tax

    Three sources quote three different prices for the identical DeepSeek model. TechTimes lists DeepSeek’s own $0.435 / $0.87. OpenRouter lists $1.218 / $2.436. Artificial Analysis measures $1.32 input and $3.96 output, with a blended rate of $0.69 per million at a 7:2:1 ratio.

    On a 500-million-token month at a typical 3:1 input-output split, that spread is real money: roughly $272 buying direct against roughly $685 through OpenRouter. Same weights, same model ID, 2.5x the invoice.

    DeepSeek has also signaled another increase is coming, with no timeline. Anyone budgeting off today’s rate should read our breakdown of the DeepSeek price increase that ended the AI price war.

    Subscription versus per-token

    Z.ai’s answer to pricing volatility is a flat plan. The GLM Coding Plan runs $18/month for Lite, $80 for Pro and $168 for Max, with weekly credit quotas of 10,000, 60,000 and 140,000 respectively.

    For a solo developer hammering an agent all day, $18 flat beats metered billing on predictability alone. For an API-backed product with variable load, per-token wins on unit economics.

    Can you actually download the weights?

    DeepSeek yes, GLM-5.3 not yet. This is the single most decisive difference between the two models and it gets buried under benchmark tables. An open-weight model you cannot download is a closed model with a press release.

    Z.ai says GLM-5.3 weights land roughly two weeks after the August 14 launch, once “safety evaluation and hardening” finish. Until then, access is API, GLM Coding Plan, or ZCode only.

    That delay is defensible given the model’s exploit-generation scores. It is still a delay, and it has three concrete consequences:

    • No air-gapped deployment. Regulated teams that cannot send code to a Chinese API endpoint are locked out entirely.
    • No cost floor. You cannot benchmark self-hosted cost per token against the API rate, which is the entire argument for open weights.
    • No fork risk protection. If Z.ai changes pricing or terms, there is no downloaded checkpoint to fall back on.

    DeepSeek’s MIT license carries none of those problems. For the economics of running weights yourself, see our analysis of what self-hosting Qwen3.8-Max really costs.

    Which model should you pick for your use case?

    Pick DeepSeek V4 Pro for long-context work, self-hosting and cost-sensitive production. Pick GLM-5.3 for security workloads and for flat-rate interactive coding. Neither beats Claude Fable 5 or GPT-5.5 on the independently verified Terminal-Bench leaderboard, so neither is the right call if raw capability is your only constraint.

    Use case Pick Why
    Self-hosted / air-gapped DeepSeek V4 Pro MIT weights available now
    Large-repo refactors DeepSeek V4 Pro 1M-token context confirmed
    High-volume API product DeepSeek V4 Pro $0.435 / $0.87 direct rate
    Vulnerability triage GLM-5.3 84.5% CyberGym, top open score
    Solo dev, predictable bill GLM-5.3 $18/month Coding Plan floor
    Long-horizon agent runs GLM-5.3 Terminal-Bench 3.0: 4.6 to 28.3
    Absolute best coding score Neither Fable 5 leads at 83.8% verified

    Is GLM-5.3 worth it in 2026?

    Yes, but narrowly. GLM-5.3 is worth paying for if you are doing security work or want a fixed monthly bill. It is not worth waiting for if you need weights on your own GPUs this quarter, and it is not the best coder available at any price.

    Z.ai calls it “the strongest open-weights coder on the market.” Its own numbers complicate that. On Terminal-Bench 3.0, GLM-5.3’s 28.3 trails Claude Fable 5 at 33.7 and GPT-5.6 Sol at 34.6. On DeepSWE v1.1, its 66.9 trails Kimi K3 at 67.5 and Fable 5 at 69.7.

    The one benchmark where GLM-5.3 leads outright is Z.ai Code Bench — Z.ai’s own benchmark. It scores 31.4% at roughly 50,000 tokens per task against Claude Opus 4.8 at 29.5% using 120,000 tokens. Efficient, and self-graded.

    Efficiency is the real story there. Beating a frontier closed model while spending 58% fewer tokens per task is a genuine margin advantage — if the benchmark holds up externally.

    Frequently asked questions

    Is DeepSeek V4 Pro open source?

    The weights are MIT-licensed and available, which permits commercial use, modification and redistribution. Training data and code are not released, so it is open-weight rather than fully open-source.

    When do GLM-5.3 weights release?

    Z.ai said roughly two weeks after the August 14, 2026 launch, pending safety evaluation and hardening. That points to late August. No license has been confirmed.

    What is the cheapest way to run DeepSeek V4 Pro?

    Direct through DeepSeek’s API at $0.435 per million input tokens on a cache miss and $0.87 output. Cache hits drop input to $0.003625. Marketplace routing costs roughly 2.5x more.

    Does GLM-5.3 beat Claude on coding?

    No, on the numbers Z.ai published. GLM-5.3 scores 28.3 on Terminal-Bench 3.0 against Claude Fable 5’s 33.7. It leads only on Z.ai’s internal Code Bench and on CyberGym.

    Why do Terminal-Bench scores differ between sources?

    Because the harness and agent scaffold change the result. Claude Code with Fable 5 scores 83.8% on the official 2.1 board, while Artificial Analysis’s own v2.1 run puts GPT-5.6 Sol at 89.5%. Only compare scores measured by the same evaluator.

    Which model has the bigger context window?

    DeepSeek V4 Pro, at 1,048,576 tokens confirmed by both OpenRouter and Artificial Analysis. Z.ai has not disclosed GLM-5.3’s context length.

    Is GLM-5.3’s cyber capability a risk?

    Z.ai treated it as one, holding weights for safety hardening after ExploitBench scores doubled to 54.4%. The model reportedly found 1,097 critical or high-severity vulnerabilities across 269 open-source projects.

    The bottom line

    Buy DeepSeek V4 Pro. Today, direct from DeepSeek, at $0.435 in and $0.87 out.

    It has the weights you can actually download under MIT, the 1M-token context you can actually verify, and a price roughly 3x below what the same model costs through a marketplace. Third place out of 1,075 models on Artificial Analysis’s Intelligence Index is enough capability for the overwhelming majority of production coding work.

    Buy GLM-5.3 in exactly two situations: your workload is security-focused, where 84.5% on CyberGym is the best open-weight number published; or you want a $18-to-$168 monthly ceiling instead of metered billing that DeepSeek has already warned will rise again.

    And discount both vendors’ headline coding claims. DeepSeek’s 87.9 on Terminal-Bench 2.1 would top the world leaderboard, and DeepSeek is not on that leaderboard. Until a third party runs the harness, those are sales figures, not results. For a comparison where the numbers were independently checked, see our breakdown of cost per coding point across Gemini 3.7 Flash and Claude Sonnet 5, and our look at Meta’s 4x cheaper coding model.

    Sources

  • Qwen3.8-Max Open Weights vs API: The Real Cost of Self-Hosting

    Qwen3.8-Max open weights are the biggest open release of 2026 — and the wrong choice for almost everyone. The 4.89 TB checkpoint is text-only, ships under a custom license with a $50 million revenue gate, and only beats the $2/$6 API somewhere north of six billion tokens a month. Below that, rent. Above it, talk to Alibaba’s lawyers first.

    Alibaba did something no Western lab has done this year: it put a 2.4-trillion-parameter frontier model on Hugging Face and told everyone to help themselves.

    Then it attached a license that quietly taxes anyone who succeeds with it.

    This is the deep dive on what the Qwen3.8-Max open weights actually cost to run, how they compare to the other trillion-scale open models, and the exact point where downloading beats paying.

    What exactly did Alibaba release with the Qwen3.8-Max open weights?

    Alibaba published the checkpoint to Hugging Face as Qwen/Qwen3.8-2.4T-A95B on August 8, 2026 — 224 files totaling roughly 4.89 TB, with an FP8 sibling repo, according to Digital Applied’s release checklist. It is a sparse Mixture-of-Experts model: 2.4 trillion total parameters, 95 billion active per token.

    That activation rate — about 4% — is the whole trick. You pay for a trillion-scale model’s quality while doing inference math on something closer to a 95B dense model.

    The announcement moved real money. Alibaba stock jumped 7% in Hong Kong and 4.5% on the NYSE on the news, Forkast reported.

    What’s missing compared to the hosted API

    The download is not the product Alibaba sells. The open checkpoint is text-only and thinking-mode only — no vision, and not the native 1M-token context the paid Qwen3.8-Max API advertises.

    Developers noticed within hours. The top thread on the model’s Hugging Face discussion board is titled “Huge disappointment,” pointing out that Alibaba’s launch post gave no hint the weights would be a stripped build. The promised smaller Qwen3.8-27B checkpoint still has not shipped.

    • In the download: 2.4T/95B MoE, text in, text out, thinking mode.
    • API only: vision and video input, 1M-token context (991K effective input cap), implicit prompt caching.
    • Still missing: the 27B variant, official deployment guidance, day-one community quantizations.

    How much does the Qwen3.8-Max API cost?

    List pricing is $2.00 per million input tokens and $6.00 per million output tokens, with cached input at $0.25 per million. That is the number every self-hosting calculation has to beat, and it is aggressive for a model in this weight class.

    Those figures are confirmed on OpenRouter’s Qwen3.8 Max listing, which also shows the 1M-token context window and a 131,072-token maximum output. Digital Applied notes the same rates were absent from Alibaba’s own Model Studio pricing page at launch — a reminder to check before you budget.

    For context on how fast this market moves: DeepSeek raised its own API prices by up to 1,100% overnight this month, which we covered in our breakdown of the collapsing AI price war. Cheap inference is not a permanent condition.

    What does self-hosting Qwen3.8-Max open weights actually cost?

    Far more than most teams assume. The FP8 checkpoint is roughly half the 4.89 TB full-precision drop — call it 2.4 TB of weights before you allocate a single byte to KV cache. That does not fit on one eight-GPU node once you leave room for long contexts.

    Start with the hardware rate. GetDeploying tracks 28 cloud providers offering NVIDIA B200 capacity: $3.35/hr at the cheapest reserved rate, $7.00/hr average on-demand, $3.83/hr average spot.

    The break-even math

    Take the friendliest possible case — a single eight-GPU B200 node at the cheapest reserved rate of $3.35/hr. That is $26.80/hr, or about $19,600 per month running continuously.

    Now blend the API price. At a 3:1 input-to-output ratio, Qwen3.8-Max costs $3.00 per million tokens blended. Divide $19,600 by $3.00 and you get roughly 6.5 billion tokens per month before the box is cheaper than the API.

    That is about 215 million tokens a day, every day, with zero idle time.

    ScenarioHourly rateMonthly cost (730 hrs)Break-even vs API
    1 node, reserved ($3.35/GPU-hr)$26.80~$19,600~6.5B tokens/mo
    1 node, on-demand avg ($7.00/GPU-hr)$56.00~$40,900~13.6B tokens/mo
    2 nodes, reserved (realistic minimum)$53.60~$39,100~13B tokens/mo
    2 nodes, on-demand avg$112.00~$81,800~27B tokens/mo
    GPU rates via GetDeploying; monthly figures and break-even points are our calculation at a $3.00/Mtok blended API price. Excludes engineering salaries, networking, and failed-run overhead.

    And that table is generous. It assumes 100% utilization, no redundancy node, and no one on payroll keeping the cluster alive. Add a single infrastructure engineer and the real break-even moves past 20 billion tokens a month.

    Does the Qwen3.8-Max license let you build a business on it?

    Only up to a point — and the point is $50 million. Alibaba abandoned the Apache 2.0 license used for earlier Qwen generations in favor of a bespoke qwen3.8-max license that forces large commercial users into a separate negotiation.

    Per Forkast’s analysis, any business operating as Model-as-a-Service or an “AI Work Assistant” with aggregate revenue above $50 million in any consecutive 12-month period must negotiate a commercial license. MaaS is defined broadly: any third-party access to inference or fine-tuning where the provider controls inputs or parameters.

    Read that structure carefully. It is a safe harbor for startups and a toll booth for anyone who scales.

    The strategic logic is obvious once you see it. Alibaba wants the distribution that open weights buy, without letting a competing inference layer get rich on top of its research budget. It is platform protection dressed as generosity — the same instinct behind Anthropic’s spending we analyzed in its $2 trillion valuation story, pointed a different direction.

    How do Qwen3.8-Max open weights compare to Kimi K3, DeepSeek V4 Pro and GLM-5.2?

    Qwen has the most parameters and the most restrictive license. DeepSeek V4 Pro has the best coding scores and by far the cheapest API. GLM-5.2 is the smallest and easiest to actually serve. If licensing matters to you, Qwen is the weakest of the four.

    ModelParams (total / active)LicenseAPI price (in / out per Mtok)Headline benchmark
    Qwen3.8-Max2.4T / 95BCustom, $50M revenue gate$2.00 / $6.00GPQA Diamond 92.6
    Kimi K32.8T / not disclosedModified MIT$3.00 / $15.00GPQA-Diamond 93.5
    DeepSeek V4 Pro1.6T / 49BMIT$0.435 / $0.87SWE-bench 80.6%
    GLM-5.2744B / ~40BMIT$1.40 / $4.40SWE-bench Pro 62.1
    Qwen figures via Digital Applied and OpenRouter; K3, V4 Pro and GLM-5.2 via MarkTechPost. Benchmark numbers come from different harnesses and are not directly comparable.

    Where Qwen wins, and where it loses badly

    Qwen3.8-Max leads on agentic and research tasks: 86.1 on OSWorld-Verified and 93.0 on PaperBench, ahead of GPT-5.6 Sol’s 90.5, per Digital Applied’s benchmark roundup.

    It loses on code. Qwen scores 67.7 on SWE-bench Pro against Fable 5’s 80.0, and 56.6 on DeepSWE 1.1 against GPT-5.6 Sol’s 73.0. DeepSeek V4 Pro’s 80.6% on SWE-bench beats Qwen outright while costing roughly a fifth as much on the API.

    The cost gap is the story. MarkTechPost’s Artificial Analysis blended cost-per-task figures put DeepSeek V4 Pro at $0.04, GLM-5.2 at $0.32 and Kimi K3 at $0.94. Paying 20x for a few benchmark points is a decision, not a default — the same trap we flagged in Gemini 3.7 Flash vs Claude Sonnet 5.

    Which model should you actually run in 2026?

    Match the model to the constraint that is actually binding you — license risk, serving budget, or raw capability. For most teams under 10 billion tokens a month, the answer is an API, and it probably is not Qwen’s.

    Your situationBest choiceWhy
    Under 5B tokens/monthQwen3.8-Max API$2/$6 with $0.25 cached input beats any cluster you can rent
    Coding agents at scaleDeepSeek V4 Pro80.6% SWE-bench, MIT license, $0.435/$0.87
    Air-gapped or regulated deploymentGLM-5.2744B/40B is the only one that fits comfortably on one node
    MaaS provider above $50M revenueAnything MIT-licensedQwen’s license forces a negotiation you will lose
    Research, agents, long-horizon tasksQwen3.8-Max APIOSWorld 86.1 and PaperBench 93.0 lead the field
    Above 20B tokens/month, MIT requiredSelf-host DeepSeek V4 Pro49B active params serve far cheaper than Qwen’s 95B

    Frequently asked questions about Qwen3.8-Max open weights

    Are the Qwen3.8-Max open weights free?

    Free to download and free to use commercially below $50 million in annual revenue. Above that threshold, Model-as-a-Service and AI assistant providers must negotiate a separate commercial license with Alibaba.

    Is the open checkpoint the same model as the Qwen3.8-Max API?

    No. The open weights are text-only and thinking-mode only, without the vision input and native 1M-token context that the hosted API provides.

    How much hardware do I need to run Qwen3.8-Max?

    The full-precision release is roughly 4.89 TB across 224 files, with an FP8 variant at about half that. Plan for multiple eight-GPU nodes, not one.

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

    Not for coding. DeepSeek V4 Pro scores 80.6% on SWE-bench versus Qwen’s 67.7 on SWE-bench Pro, ships under MIT, and costs roughly a fifth as much per token.

    Why did Alibaba drop Apache 2.0?

    Commercial strategy. Apache 2.0 would have let rival inference providers build businesses on Alibaba’s research at zero cost. The revenue gate keeps distribution while capturing the upside at scale.

    When does self-hosting Qwen3.8-Max become cheaper than the API?

    Around 6.5 billion tokens a month in the best case, and realistically past 13 billion once you run two nodes. Add engineering headcount and the crossover pushes past 20 billion.

    Did the Qwen3.8-27B model ever ship?

    Not as of mid-August 2026. The smaller checkpoint was announced alongside Max but has not appeared on Hugging Face.

    The bottom line

    Use the Qwen3.8-Max API if you are under roughly five billion tokens a month and you need agentic or research performance — OSWorld 86.1 and PaperBench 93.0 are genuinely class-leading, and $2/$6 is fair for that tier.

    Do not self-host it. The break-even sits above 13 billion tokens a month at realistic node counts, and the model you would be hosting is the stripped text-only build, not the one that posts those benchmark numbers.

    If you are building anything you intend to sell inference on, pick DeepSeek V4 Pro or GLM-5.2 instead. MIT licensing costs nothing at $50 million in revenue. Alibaba’s license costs you a negotiation with a company that also competes with you.

    The open-weights headline was real. The gift was not. Qwen3.8-Max is the most capable model you can legally download this month and the one with the most expensive fine print — and for once, both halves of that sentence matter equally. For more on how speed and cost trade off at the frontier, see our analysis of what 14x inference speed actually costs.

    Sources