Open Source Moved the Needle This Week: GLM-5.1 Beat GPT-5.4, Gemma 4 Went Apache 2.0

The headlines this week were all about Claude Mythos and GPT-5.4-Cyber, closed-door vetted-access security dramas. Meanwhile, open-source made two moves that deserved more attention.

1. April 7, 2026, Z.ai (formerly Zhipu AI) released GLM-5.1. SWE-bench Pro: 58.4. That’s past GPT-5.4 at 57.7 and Claude Opus 4.6 at 57.3. MIT license. Weights downloadable. Runs locally.

2. In the same week, Google moved Gemma 4’s license to Apache 2.0.

These look like parallel stories. Stacked together, they’re the same story: open source unblocked its two largest bottlenecks (capability parity and commercial licensing) in the same week.

The GLM-5.1 Facts

Pulling from buildfastwithai’s review, zenvanriel’s technical breakdown, and MindStudio’s overview:

Architecture: 754B parameter Mixture-of-Experts, targeted at agentic engineering and long-horizon tasks License: MIT — commercial use, modification, distillation, no royalties Distribution: Weights public on Hugging Face Compute: Reportedly trained without any US-origin chips Business: Z.ai went public on the Hong Kong Stock Exchange in January 2026, raising roughly $558 million. The first publicly listed AI foundation-model company.

The capability numbers:

Benchmark	GLM-5.1	GPT-5.4	Claude Opus 4.6
SWE-bench Pro	58.4	57.7	57.3

The margin is small. The direction matters more. For the first time, an open-weight model has simultaneously beaten both OpenAI’s and Anthropic’s flagships on a mainstream coding benchmark.

And one measured characteristic is more interesting than the headline score: GLM-5.1 can run a plan → execute → test → fix → optimize loop autonomously for up to 8 hours. That’s not code completion. That’s actual agentic work.

Gemma 4 Going Apache 2.0 Is the Bigger Signal

I wrote in the earlier Gemma 4 piece that “Apache 2.0 might be the most important single line in the announcement.” This week validates that read.

Google’s Open Source Blog formally moved Gemma 4 off the Gemma license (which carries commercial limitations) to Apache 2.0. That essentially matches GLM-5.1’s MIT. Between them, these are the two most permissive licenses in common industry use.

What that changes:

• Legal friction drops. The thing that stops a model from hitting production is often the license boundary. Apache 2.0 and MIT almost don’t have that problem
• Fine-tuning and distillation tools will support them first. Tools like Unsloth, Axolotl, and LLaMA Factory prioritize MIT / Apache 2.0 over custom licenses
• Internal enterprise PoCs become easier to start. “Let’s try the open-source version” used to get killed by legal. That blocker weakens

Licensing vs. Capability — Which One Matters More?

Old argument. The dominant view for the last two years has been “closed source runs 6–12 months ahead.” That was true in early 2024 and started weakening through the second half of 2025. After this week, I think the framing is broken.

Real procurement decisions are almost never “which score is higher.” They’re:

1. Is the capability good enough for my use case? (For a lot of use cases, GPT-5.4 and Claude Opus are overkill)
2. Can I legally deploy this into my product?
3. How locked in am I a year from now?
4. If this vendor folds, does my infra fold with it?

Apache 2.0 and MIT answer questions 2, 3, and 4 better than any closed API. The thing that used to block adoption was question 1: open weights genuinely weren’t good enough. That bar just got cleared.

None of this means Anthropic or OpenAI are suddenly in trouble. Their product layer (Claude Code, ChatGPT Enterprise, things like Mythos) still has real value in security, compliance, and integration. But the “I just need a good-enough general LLM” market is running out of reasons to pay per-token to closed labs.

The details from the model cards

GLM-5.1. Z.ai’s model card describes it as a flagship for “agentic engineering” and publishes it under the MIT licence. Its own benchmark table puts GLM-5.1 at 58.4 on SWE-Bench Pro, ahead of GPT-5.4 (57.7) and Claude Opus 4.6 (57.3) in that table, and at 63.5 on Terminal-Bench 2.0 with the Terminus-2 harness. The same table shows it behind closed models elsewhere: Opus 4.6 leads on NL2Repo (49.8 vs 42.7), and Gemini 3.1 Pro and GPT-5.4 lead on Humanity’s Last Exam without tools. Z.ai’s main claim is about long tasks: that the model keeps improving over hundreds of rounds and thousands of tool calls instead of plateauing early. The card lists SGLang (v0.5.10+) and vLLM (v0.19.0+) for local serving. These are vendor-reported numbers, so independent results are worth waiting for.

Gemma 4. Google’s model card lists Apache 2.0 as the licence and covers a range of sizes: small “effective” models (E2B and E4B) for phones and laptops, a 12B model, a 26B mixture-of-experts model that activates about 3.8B parameters per token, and a 31B dense model. Small models have a 128K-token context window and larger ones 256K. All accept text and images, and the smaller ones also take audio. Google also highlights configurable thinking modes, native function calling and native system-prompt support.

Which one fits which job

• Self-hosted coding agents on serious hardware: GLM-5.1 is the more relevant model. It targets long agentic coding sessions, and the MIT licence imposes very few conditions.
• On-device or low-cost deployment: Gemma 4’s small and MoE variants are built for laptops, phones and single GPUs, where GLM-5.1’s size is impractical.
• Products that need clear legal terms: both licences are standard and widely understood, which matters more to legal teams than a few benchmark points.
• Highest accuracy regardless of cost: closed frontier models still lead on several of the benchmarks in Z.ai’s own table, so test before switching.

What to Watch

If this affects decisions on your side, a few things to track over the next couple of months:

• Z.ai’s next training run. 754B MoE got to 58.4 this time. The real test is H2 2026, when GPT-6, Claude 5, and Gemini 4 are all expected. Whether open source holds its position against those is the real question
• Government constraints on Apache 2.0 / MIT models. US-China relations, export controls, and AI safety regulation could flip this trend in 2026. GLM-5.1’s “no US chips” training detail has geopolitical weight built in
• Hugging Face’s role. As open source becomes a production option, Hugging Face shifts from “GitHub for researchers” to “enterprise AI procurement portal.” Whether that transition goes smoothly depends on their enterprise support
• Fine-tuning tool explosion. Apache 2.0 opens the path for fine-tuning platforms to ship genuinely different products from cloud vendors

The closed-source drama was loud this week. But open-source submitted the more durable answer. SWE-bench Pro 58.4 is not the main point. MIT and Apache 2.0 are.