Coding Horizon

The best coding AI for 4GB VRAM has a hidden rival

Every figure, size, date and benchmark score this video puts on screen, traced to the place that published it.

Weight file sizes are the Q4_K_M build unless stated otherwise, because that is the quantisation a runner offers first and the one a viewer will actually download.


Liquid AI LFM2.5 2.6B

Released 4 August 2026.

Claim Value Source
Q4_K_M weight file 1.67 GB LiquidAI/LFM2.5-2.6B-GGUF
Q4_0 weight file 1.59 GB same
Throughput, AMD Ryzen AI Max+ 395 113 tok/s Liquid AI launch post
Throughput, Apple M5 Max 220 tok/s same

The model card gives both figures but names only “an AMD Ryzen CPU”. The specific part, Ryzen AI Max+ 395, appears in Liquid AI’s own launch post, which is what the beat credits. | Context window | 128K | same |

Recommended uses, from the model card: “We recommend using it for agentic workloads, tool use, data extraction, RAG, and long-context workflows.”

The limit the video quotes, also from the model card: “It is not recommended for agentic coding and knowledge-heavy tasks.”

Both throughput figures are the publisher’s own, measured on the machines named. Neither describes a GTX 1650.

Sources:


Microsoft Phi 4 mini instruct

Claim Value Source
Parameters 3.8B microsoft/Phi-4-mini-instruct
Context window 128K same
Q4_K_M weight file 2.49 GB bartowski/microsoft_Phi-4-mini-instruct-GGUF

Microsoft’s own summary of what the model is for names “Strong reasoning (especially math and logic)”, which is the emphasis the video refers to.

Sources:


Nanbeige 4.1 3B

Claim Value Source
LiveCodeBench v6 76.9 Nanbeige/Nanbeige4.1-3B
Qwen3-30B-A3B-2507, same table, same benchmark 66.0 same
Community q4, Q4_0 2.33 GB mradermacher/Nanbeige4.1-3B-GGUF
Community q4, Q4_K_S 2.34 GB same
Community q4, IQ4_XS 2.24 GB same
Community q4, Q4_K_M (the default) 2.44 GB same

There is no official Nanbeige GGUF; every q4 above is a community build. The script says “around 2.3”, which is true of Q4_0, Q4_K_S and IQ4_XS but NOT of Q4_K_M, the build most runners pull by default. The picture therefore names the quant it is showing rather than printing a bare “2.3 GB”, because an unlabelled figure would read as the default one.

The comparison the video makes is the publisher’s own table: the 3B model’s 76.9 sits above Qwen3-30B-A3B-2507’s 66.0, a model with thirty billion total parameters. The rest of that column, for context: Qwen3-4B-2507 57.4, Qwen3-8B 49.4, Qwen3-14B 55.9, Qwen3-32B 55.7.

There is no official GGUF release, so the weight file sizes are community conversions: IQ4_XS 2.3 GB, Q4_K_S 2.4 GB, Q4_K_M 2.5 GB.

Sources:


IBM Granite 4.2 3B

Released 25 August 2026, Apache 2.0.

Claim Value Source
Q4_K_M weight file 2.24 GB ibm-granite/granite-4.2-3b-GGUF
LiveCodeBench v6 69.71 ibm-granite/granite-4.2-3b
Context window 128K same

The three thinking modes the video describes are settings on the chat template:

The blank entries the video points at are real. On IBM’s own benchmark table the 3B model has no published result for SWE-bench Verified, SWE-bench Multilingual, SWE-bench Pro, Terminal-Bench 2.1, BirdBench or GDPval. The larger models in the same family do carry results in those rows.

Sources:


Google Gemma 4 E2B

Claim Value Source
What the E stands for effective parameters google/gemma-4-E2B-it
Parameters, with embeddings 5.1B same
Effective parameters, without 2.3B same
LiveCodeBench v6 44.0 same
Q4_K_M weight file 3.11 GB unsloth/gemma-4-E2B-it-GGUF
Vision projector, F16 0.99 GB same
Modalities text, image, audio google/gemma-4-E2B-it

The gap the video is pointing at is between the name and the download. The name counts 2.3B effective parameters; the file has to store 5.1B, because per layer embedding tables are looked up rather than multiplied through, and they are still bytes on disk. The vision projector is a separate file on top of the 3.11 GB.

Sources:


Qwen3.5 4B

Claim Value Source
Parameters 4B Qwen/Qwen3.5-4B
LiveCodeBench v6 55.8 same
Q4_K_M weight file 2.74 GB unsloth/Qwen3.5-4B-GGUF
Vision projector, F16 672 MB same
Image input yes Qwen/Qwen3.5-4B

Thinking can be turned off through the chat template: "enable_thinking": False passed in chat_template_kwargs. Qwen notes that Qwen3.5 does not support the older soft switch.

Sources:


The benchmark itself

LiveCodeBench evaluates a model by running its generated solutions against test cases for programming problems, rather than comparing them to a reference answer as text. Version 6 is the contamination controlled window used by every score above.

The six scores this video shows, all version 6, all from the publisher of the model in question:

Model LiveCodeBench v6
Nanbeige 4.1 3B 76.9
Granite 4.2 3B 69.71
Qwen3-30B-A3B-2507 66.0
Qwen3.5 4B 55.8
Gemma 4 E2B 44.0

They come from different labs running their own evaluations with their own settings, and they measure the full precision weights, not the four bit files that fit on a 4 GB card.

Source: https://livecodebench.github.io/


The hardware

Card Memory Bandwidth Source
GTX 1650, GDDR5 4 GB, 128 bit 128 GB/s NVIDIA GTX 16 series
GTX 1050 Ti 4 GB, 128 bit 112 GB/s PC Gamer, GTX 1650 review
RTX 3050 Laptop, 4 GB 4 GB GDDR6, 128 bit 192 GB/s Notebookcheck

The RTX 3050 laptop part is the newer card and moves memory half again as fast as a GTX 1650, and it still has the same four gigabytes to fit a model into. Bandwidth describes how fast memory moves data; capacity decides what loads at all.

Sources:


Where the model actually ran

ollama ps reports placement in its PROCESSOR column: 100% GPU when the whole model is on the card, 100% CPU when no usable GPU was found, and a split such as 48%/52% CPU/GPU when the weights and the KV cache together are larger than the free VRAM. A split is the runner doing its job, not a fault, and it is the first thing to read before blaming the model for being slow.

Sources:


Caveats


Verified before publish

Every figure this video puts on screen was re-checked against a primary source after the script’s final rewrite. Confirmed exactly as shown: Liquid’s 1.67 GB Q4_K_M and the sentence ruling out agentic coding; phi 4 mini’s 3.8B and 2.49 GB; Nanbeige’s 76.9 against Qwen3-30B-A3B-2507’s 66.0 in its own table; Granite 4.2 3B’s August 25 release, 2.24 GB official Q4_K_M and 69.71; Gemma 4 E2B’s effective-parameter naming, 5.1B with embeddings, 44.0 and Unsloth’s 3.11 GB; Qwen3.5 4B’s 55.8, 2.74 GB, image input and disableable thinking; the GTX 1650 GDDR5 at 128 GB/s and the GTX 1050 Ti at 112 GB/s, both from NVIDIA’s own spec tables.

Three things were corrected rather than confirmed:

Not checked