Open Weights Ship Three Frontier Models in One Week

A year ago, Paul Graham declared that tuning open-weight models was "a waste of time." The consensus held: proprietary labs would leapfrog anything the open community could build, rendering fine-tuning futile. That consensus died this week. Between August 12 and 14, three Chinese labs — Zhipu (Z.ai), Alibaba (Qwen), and DeepSeek — each shipped frontier-class models with open weights or imminent weight releases. GLM-5.3 took the #1 slot on Hacker News with 924 points. Qwen3.8 became the first Max-class model downloadable from Hugging Face. DeepSeek V4 Pro went GA alongside an MIT-licensed agent harness. This is not three coincidental releases. This is a coordinated industrial response to the proprietary moat thesis — and it is shipping product.

📖 Read the full version with charts and embedded sources on ComputeLeap →

View original post on X →

The through-line matters more than any single model. Open weights are no longer a philosophical position or a research convenience. They are becoming a market strategy, backed by state-aligned capital, executed by labs that have closed the benchmark gap to single-digit percentages, and delivered at a fraction of the cost of proprietary alternatives.

GLM-5.3: The Model That Found Bugs Its Creators Did Not Expect

Zhipu's GLM-5.3, released August 14, is the headline act — and deservedly so. Built on the exact same base model as GLM-5.2, every improvement came from scaled-up post-training alone. No new pretraining run, no architecture change. Just reinforcement learning and training-environment design pushed further than the team anticipated.

The results are striking. On coding benchmarks, GLM-5.3 hit 28.3 on Terminal-Bench 3.0 (up from GLM-5.2's 4.6) and 66.9 on DeepSWE v1.1 (up from 46.2). On Agents' Last Exam it reached 28.5, up from 23.8. At maximum thinking intensity on Z.ai's internal benchmark, it hit 34.5% using roughly 75,000 output tokens.

But the real story is cybersecurity. GLM-5.3 scored 84.5% on CyberGym — first place, ahead of Anthropic's Mythos 5 (83.8%) and OpenAI's GPT-5.6 Sol (83.6%). On ExploitBench, it hit 54.4%, more than doubling GLM-5.2's 24.4%. The model developed what Z.ai calls "emergent cyber capability" — multi-step exploit-chain reasoning that the company says it did not plan for and did not explicitly train.

View on Hacker News →

The vulnerability discovery ledger tells the scale. Z.ai reports 2,436 vulnerabilities found across 269 open-source projects, including flaws dating back 45 years. By severity: 107 Critical, 990 High, 1,286 Medium, 53 Low. Fifty-three CVEs have been publicly assigned; 2,383 remain under coordinated embargo at cvd.z.ai. The disclosed bugs include a Linux kernel 6lowpan use-after-free (CVE-2026-64452), a WebKit/Safari memory-handling flaw (CVE-2026-43663), and a FreeBSD ptrace validation bug (CVE-2026-45253).

One HN commenter captured the practitioner reaction: they bought the $18 GLM subscription, set it up with a Claude Code harness for security research, and "bumped to $80 plan almost immediately." The model agreed to perform proper red-team analysis and executed it without hesitation.

View original post on X →

Jeremy Howard, co-founder of fast.ai, has been direct in his assessment of the GLM line, calling GLM-5.2 "at least as good as Opus 4.8 and GPT 5.5" while being "super fast, inexpensive, not too verbose." GLM-5.3 builds on that foundation with post-training gains that nobody in the open-weight space was achieving a year ago.

Qwen3.8: The First Downloadable Max-Class Model

While GLM-5.3 was dominating HN, Alibaba quietly dropped something arguably more significant for the open-weight ecosystem. On August 12, Qwen released the weights for Qwen3.8-2.4T-A95B — the first Max-class model available as a downloadable artifact on Hugging Face.

The numbers: 2.4 trillion total parameters, roughly 95 billion active parameters, 512 experts (10 routed + 1 shared per layer), an 8192 hidden dimension across 92 layers. Native context window of 262,144 tokens, extensible to over 1 million. This is not a small model you run on a laptop. It is a full frontier-class MoE that organizations with GPU clusters can now self-host.

The benchmark profile confirms the ambition. GPQA Diamond: 92.6. SWE-bench Pro: 67.7. PaperBench: 93.0. Terminal-Bench 2.1: 86.6. These are numbers that compete directly with the best proprietary offerings on key developer-relevant tasks.

View original post on X →

Clement Delangue, CEO of Hugging Face, amplified the release with a single-word endorsement. The enthusiasm is justified. Until now, Max-class Qwen models were API-only. Making the weights downloadable changes the sovereignty equation for any team that cares about data residency, fine-tuning control, or simply not being locked into a single provider's pricing schedule.

The model requires thinking mode for all interactions and is text-only — no multimodal capabilities in this release. But for coding, research, and agent-driven workloads, it is the most capable open-weight model you can download today. For more context on running open models locally, see our guide to running Qwen3 35B on a local Mac.

DeepSeek V4 Pro: The Model Plus the Toolkit

DeepSeek's contribution to this week's open-weight barrage came on August 13, with the GA release of DeepSeek V4 Pro 0813 and the simultaneous launch of DeepSeek Harness v0.1.

The model improvements are substantial. From the April preview to GA, DeepSWE jumped from 12.8 to 62.7 (+49.9 points), CyberGym from 52.7 to 83.3 (+30.6), NL2Repo from 38.5 to 61.5, and Terminal-Bench 2.1 from 72.1 to 87.9. The weights are MIT-licensed on Hugging Face — roughly 893GB across 67 safetensors files, built on a 1.6T total / 49B active parameter architecture with a new DSpark speculative decoding module.

View on Hacker News →

As one HN commenter noted: "Competitive with Opus 4.8 but weaker than Sol or Fable. About 20x cheaper." That price-performance ratio is the open-weight value proposition in a sentence. For a deeper look at how DeepSeek V4 compares to other frontier models, see our earlier comparison.

But the real differentiator is DeepSeek Harness. Released under MIT license, it is a provider-agnostic agent framework built on a radical modularity principle: everything is a plugin. Inference, tools, session state, the agent loop itself — all replaceable. It supports native OpenAI Responses API format, integrates with Codex, and offers three thinking effort levels.

This is significant because it addresses the harness problem directly. The competitive advantage is shifting from the model to the scaffolding around it. DeepSeek is not just releasing weights — it is releasing the complete development environment.

What the Market Is Pricing

The Polymarket prediction markets provide a useful reality check on how the betting public is processing this week's releases.

"Second-best AI Lab end of August" prices Alibaba at 55% and Google at 37%. That is a remarkable shift — Alibaba, maker of Qwen, is now the market favorite for the #2 position behind Anthropic. "Third-Best Chinese AI Company end of August" prices Z.ai (Zhipu, maker of GLM) at 52%, ahead of Baidu at 28% and Moonshot at 11%.

These are not sentiment polls. These are real-money bets from traders with financial incentive to be right. And they are telling us that the market considers open-weight Chinese labs to be the primary challengers to proprietary Western providers.

View original post on X →

The broader context reinforces the signal. In the same week, Meta shipped Muse Glimmer — a 30B open-weight agentic model under Apache 2.0 that Alexandr Wang called "personal superintelligence available to everyone." The open-weight wave is not limited to Chinese labs. It is a cross-continental consensus that weights should be downloadable, forkable, and self-hostable. For our coverage of the broader open-weight frontier including Kimi K3 and Inkling, see this analysis.

The Convergence Is the Story

Any one of these releases would be notable. The three arriving in the same week is the actual signal.

Consider the structural picture. Zhipu proved that post-training alone can push an existing base model past proprietary frontier benchmarks on a specialized surface (cybersecurity). Alibaba demonstrated that Max-class model weights can be made available for download — breaking the API-only barrier that kept the most capable models locked behind provider control. DeepSeek showed that the value chain extends beyond weights to tooling, with an MIT-licensed harness that lets any team build agentic systems without vendor lock-in.

Together, these three releases form a coherent thesis: the closed-model moat is narrowing on benchmarks, collapsing on price, and now being flanked on tooling. Our earlier deep-dive into GLM-5.2's post-training factory predicted this trajectory — the post-training scaling paradigm Z.ai pioneered with slime has now been validated by GLM-5.3's results.

⚠️ Contrarian Corner: Are These Weights Really "Open"?

Before declaring victory for open weights, note three inconvenient facts. First, GLM-5.3's weights are not yet released — Z.ai is holding them for "safety evaluation and hardening" tied to the model's cyber capabilities. When a lab delays its own open-weight release for safety, the "open" label is marketing until the bits ship. Second, these models still trail Sol and Fable 5 on general coding benchmarks — CyberGym is a narrow surface, and ExploitBench results are self-reported. Third, Polymarket still prices Anthropic as the #1 lab overall. The counteroffensive is real, but the war is not won.

What This Means for You

If you are building agentic coding systems, evaluating model providers, or making infrastructure bets, here is the concrete takeaway:

The viable model roster just tripled. Before this week, frontier-class coding required Claude, GPT-5.6, or Gemini. Now GLM-5.3 at $18/month, Qwen3.8 for self-hosted deployment, and DeepSeek V4 Pro at roughly 1/20th the price of Opus 4.8 are all credible options for production workloads.

Benchmark your own tasks. Aggregate benchmarks obscure workload-specific performance. GLM-5.3 leads on cybersecurity but trails on general coding. Qwen3.8 excels on research and planning. DeepSeek V4 Pro is the strongest on pure value-per-token. The only way to know which model fits your pipeline is to test it.

Evaluate DeepSeek Harness. If you are currently locked into a single provider's agent framework, the MIT-licensed Harness is worth a weekend of experimentation. Its plugin architecture means you can swap models without rewriting your agent logic — a genuine hedge against provider lock-in.

Watch the weights drop. GLM-5.3's open-weight release is expected within two weeks. When it ships, any team with GPU capacity will have access to a model that beats Mythos 5 on cybersecurity benchmarks. The security implications — both defensive and offensive — deserve serious attention from every CISO and security team.

ℹ️ The Bottom Line: Paul Graham was wrong a year ago and right this week. Open-weight tuning is back — not because the community caught up in a grassroots way, but because three well-funded labs made it an industrial strategy. The counteroffensive has shipped. The question is no longer whether open weights can compete with proprietary frontier models. It is whether proprietary providers can justify their pricing premium when the gap is this small.

For more on the open-weight trajectory, see our earlier coverage of GLM-5.2's post-training breakthrough, the GLM-5.2 vs Opus comparison, and the broader open-weight frontier momentum.

Originally published at ComputeLeap