Coding Horizon

IBM Granite 4.2: Use 3B Before Claude Or GPT 5

Every figure, date, version and benchmark this video puts on screen, chased to a primary source. Where a number could not be sourced, the shot does not render it.

The release

Claim	Finding	Source
Released 25 August 2026	Confirmed	IBM Research, Granite 4.2 brings native reasoning to enterprise agents
Three sizes: 3B, 8B, 30B	Confirmed, plus quantized variants per size	ibm-granite/granite-4.2-language-models
Apache 2.0	“All Granite 4.2 Language Models are distributed under Apache 2.0 license”	ibm-granite/granite-4.2-language-models
128K native context across the family	Native context is 131,072 tokens, which is what “128K” refers to	IBM, Granite 4.2 LLMs: How They are Built
Long context extension up to 512K	Staged pre-training extends the window to 512K tokens	IBM, Granite 4.2 LLMs: How They are Built
Thinking, non thinking, and low effort thinking modes	Three modes, the third spending “a short reasoning budget on easy questions”	IBM, Granite 4.2 LLMs: How They are Built
Native tool calling	Confirmed	IBM, Granite 4.2 LLMs: How They are Built
Agentic reinforcement learning on software engineering, terminal and search	Applied to the 8B and 30B only. IBM describes “enterprise-style tasks, including software engineering, terminal-based coding, and search-driven workflows”	IBM Research, IBM on Hugging Face

Published benchmark results

All figures are IBM’s own reported results for Granite 4.2, from IBM’s write up of how the models were built. They have not been independently reproduced at the time of recording, which is why the video says so on screen rather than presenting them as settled.

Benchmark	3B	8B	30B
AIME 2025	78.33	86.67	89.17
LiveCodeBench v6	69.71	73.24	75.77
MMLU Pro	67.84	74.04	77.60
BFCL v4	52.41	50.29	61.39
GPQA	54.80	64.14	66.41
Arena Hard V2	34.96	65.19	67.93
RULER at 128K	55.30	71.41	81.38
SWE Bench Verified	not reported	47.67	57.00
SWE Bench Pro	not reported	19.11	33.29
Terminal Bench 2.1	not reported	20.56	29.24

Source: IBM, Granite 4.2 LLMs: How They are Built

Two things in that table are load bearing for the argument and are drawn as they stand:

• The 3B has no SWE Bench Verified or Terminal Bench figure at all. IBM does not report one. The video shows those cells empty rather than filling them, because the absence is the point being made.
• The 8B scores lower than the 3B on BFCL v4 (50.29 against 52.41). The video does not claim otherwise, and no shot draws BFCL as rising with size.

Deployment and platform support

The video says IBM lists support across Hugging Face, Ollama, LM Studio, vLLM, SGLang and llama.cpp through GGUF. That spans two IBM write ups and both halves check out:

Platform	Source
Hugging Face, Ollama, LM Studio, GitHub, watsonx	IBM Research, which also notes the 3B fits a laptop through Ollama or LM Studio and the 30B serves through vLLM
vLLM, SGLang, llama.cpp with GGUF conversions	IBM, Granite 4.2 LLMs: How They are Built

Ollama announced availability of all three sizes on the day of release. Source: Ollama

Llama 3.2, for the second comparison

Claim	Finding	Source
Released 25 September 2024	Confirmed	Meta, Llama 3.2: Revolutionizing edge AI and vision
1B and 3B lightweight models	Confirmed	Meta
Built for edge and mobile	Confirmed, with day one support for Qualcomm and MediaTek hardware and Arm optimisations	Meta
Pushed for summarization, rewriting, instruction following	Confirmed as the named on device use cases	Meta
“Almost two years later”	23 months between the two releases	Arithmetic on the two dates above
1B and 3B context length	128K tokens	Meta

The last row matters as a caution rather than as a claim: Llama 3.2 3B also has a 128K window, so no shot in this video draws Granite’s native context as an advantage over Llama. The context comparison on screen is against Qwen only.

Qwen3, for the third comparison

Claim	Finding	Source
Qwen3 8B is 32K native and extends to 131K with YaRN	Confirmed. original_max_position_embeddings is 32,768 and a YaRN factor of 4.0 reaches 131,072	Qwen/Qwen3-8B model card
More than 100 languages and dialects	Qwen’s own figure is 119 languages and dialects, and Qwen’s documentation also uses the “100+” phrasing	Qwen3 Technical Report, Qwen3 blog
Qwen3 4B and 8B are Apache 2.0	Confirmed on the model cards	Qwen/Qwen3-8B

Not verified

• The claim that Meta pushed Llama 3.2 3B for agentic retrieval specifically. Meta’s announcement names summarization, instruction following and rewriting, and separately describes tool calling and on device agentic applications. “Agentic retrieval” is a fair characterisation of that positioning rather than a phrase quoted from Meta, so no shot renders it as a quotation.
• Every Granite 4.2 benchmark figure above is vendor reported. No independent reproduction had been published at the time of recording. The shots that show these numbers label them as IBM’s own.
• The cost figures and call counts that appear on the cost and routing shots are illustrations of shape, not measured prices. They are never attributed to IBM, never labelled as a benchmark, and no shot presents them as a published figure.
• The share percentages on the routing ladder are likewise an illustration of how a tiered stack distributes work, not a measurement of any particular deployment.