Z.ai and IBM ship open-weight reasoning models a day apart

Two open-weight reasoning models shipped within a day of each other. Z.ai released GLM-5.3-Flash on 26 August 2026 under an MIT licence, per MarkTechPost. IBM released Granite 4.2 on 25 August 2026 under Apache 2.0, per DataNorth. Open weights means the model files themselves are published, so you can download and run them yourself. The two are aimed at almost opposite hardware, which is the part worth understanding before you pick one.

First, a naming trap. GLM-5.3 and GLM-5.3-Flash are different releases. DataNorth reports GLM-5.3 arrived on 14 August 2026 for coding and cybersecurity work, with open weights promised about two weeks later. GLM-5.3-Flash is the 26 August release, and its weights are already on Hugging Face.

GLM-5.3-Flash: large, cheap to call, hard to host

GLM-5.3-Flash is a mixture-of-experts model with 320 billion total parameters and 18 billion active per token, according to MarkTechPost. A mixture of experts only switches on a small share of its parameters for each token, so it can be large without costing large amounts of compute per token. Here it activates 8 of 288 experts.

MarkTechPost reports it is the first natively multimodal model in the GLM-5 series, taking image and video input, with a context window of 1,048,576 tokens. Z.ai calls it "the first natively multimodal model in the GLM-5 series and the cheapest capable coding model". The efficiency claims are specific: 3 times less attention compute and a 4.4 times smaller KV cache than GLM-5.2.

Pricing is where the pitch lands. MarkTechPost lists $0.15 per million input tokens, $0.03 per million cached input tokens, and $0.50 per million output tokens. Z.ai says the model "beats GLM-5.2 across benchmarks and real workloads at roughly one-tenth the price".

The catch is in the same report. Self-hosting needs substantial hardware, with the weights around 306 GiB, and its vision ability trails Gemini 3.7 Flash.

Granite 4.2: small enough to actually run

IBM went the other direction. DataNorth reports Granite 4.2 comes in 3B, 8B, and 30B parameter sizes under Apache 2.0, trained on roughly 15 trillion tokens including 1 trillion synthetic code tokens. Context is 131,072 tokens, with a claimed extension to 512K, and it supports 12 languages.

Its notable design feature is what IBM calls "a thinking switch, so one checkpoint can either reason step by step or answer directly". That matters practically: you ship one model file and choose per request whether to pay for reasoning tokens. DataNorth reports the 30B model scores 57.00 on SWE-bench Verified, a benchmark that asks a model to fix real GitHub issues.

IBM also released Granite Speech 5.0 Turbo CTC models at 470M parameters, which DataNorth reports hit 12,600 RTFx throughput on an NVIDIA H200.

Side by side

	GLM-5.3-Flash	Granite 4.2
Sizes	320B total, 18B active	3B, 8B, 30B
Licence	MIT	Apache 2.0
Context window	1,048,576 tokens	131,072 tokens
Inputs	text, image, video	text
Weights footprint	about 306 GiB	not stated
Named benchmark	Terminal-Bench 3.0 at 28.3% for GLM-5.3	SWE-bench Verified 57.00 at 30B

What this means for developers

Pick by where the model will run, not by parameter count. A 320B model at 306 GiB is a multi-GPU deployment or an API call. A 3B or 8B model is something you can put on one machine, including hardware you already own. If your requirement is offline or on-premise, the Granite sizes are the only ones on this list that make that easy, regardless of which model scores better.

Do not compare the two headline benchmark numbers against each other. Terminal-Bench 3.0 and SWE-bench Verified are different tests, so 28.3% and 57.00 are not on the same scale. The comparable number here is GLM-5.3 against GLM-5.2 on the same benchmark, where DataNorth reports a jump from 4.6% to 28.3%. That gain came from post-training alone, on an unchanged base model, which is a useful reminder that the base is often not the bottleneck.

Both licences are permissive, but they are not identical, and the difference shows up in legal review. MIT and Apache 2.0 both allow commercial use and modification. Apache 2.0 adds an explicit patent grant, which some legal teams require. Check which one your organisation has already approved before you build on either.

The pricing on GLM-5.3-Flash is the number that should change a plan. At $0.03 per million cached input tokens, a workload that re-sends the same long context repeatedly gets dramatically cheaper, which is exactly the shape of most agent and document-analysis workloads. If you rejected long-context designs on cost last year, that calculation is worth redoing. Verify the cache-hit behaviour on your own traffic first, since cached pricing only helps if your prompts actually repeat.