On-Premises AI Lab

🧪 Missing a tree in Godot? Qwen3.5-4B-GLB-3D-Coder, a tiny 4B model built just for GLB assets｜Local AI Lab

Mark Ku

November 3, 2026•29 min read

Opening

5.2k downloads, 103 likes, HF trending score 100: ccore/Qwen3.5-4B-GLB-3D-Coder is a 4B, Apache-2.0 model that ships with GGUF out of the box, and the official model card says it was trained to do exactly one thing: generate game assets for Godot editors that use GLB files. The other two items today go in completely the opposite direction, one prunes half the experts out of a 763.2B MoE to fit it on local hardware, the other is a brand-new 560B total-parameter text-only architecture. As for just how big a file "pruning half the experts" actually leaves you with, I'll get to that shortly.

This Issue's Picks

1. ccore/Qwen3.5-4B-GLB-3D-Coder: a 4B model purpose-built for Three.js scenes with a button to export GLB

Source / summary: Picked up from Hugging Face by this show's "local model radar" (https://huggingface.co/ccore/Qwen3.5-4B-GLB-3D-Coder ). The official model card labels it Early preview and positions it as "3D scene and asset generation with GLB export," with the author noting that training is still ongoing.
What it is: A 4B coding-specialized model under Apache-2.0, currently released only in GGUF form (Q8_0 and Q4_K_M). The author describes it as "a small 4B coding model trained to generate Three.js scenes and GLB-exportable assets." Note that the full Safetensors weights and the training dataset, according to the model card, won't be released until training finishes.
Where you'd use it:

(1) Indie game devs who suddenly need an asset inside Godot. You're building a small game, the scene is still missing a tree, a mook enemy, a moving platform, and your outsourced art hasn't come back yet, but you want to finish blocking out the level today. The official model card's stated use case is exactly "generating game assets for Godot editors that use GLB files." A separate third-party description mentions that it outputs a standalone Three.js HTML file with an "Export GLB" button, so the developer just clicks it and drags the asset into the project. That third-party lead comes from issue #906 in a GitHub project called haruspex, but the selection brief only recorded the issue number without the full repo name or link, so it's less verifiable than the official model card. Treat it as a third-party report.

(2) Frontend / WebGL engineers prototyping interactive scenes. A client wants a 3D product showcase page and you want to throw out three or four scene drafts to discuss direction, but the project content can't go to the cloud. The 4B + GGUF combo can sit on local llama.cpp as a scene drafting machine, with code and requirements never leaving the box.

(3) Anyone building a "big model dispatches to small models" architecture. That haruspex issue treats it as the first use case for a dedicated specialist model, saying it's "too small for general coding, but good at this one thing." If you're self-hosting an agent toolchain and want the main model to farm out a narrow task like "generate 3D assets" to a resident 4B model, this is an off-the-shelf piece.
Can you run it: The official model card lists no VRAM threshold, only the downloadable quantizations: Q8_0 and Q4_K_M GGUFs, plus an optional unmodified stock Qwen3.5-4B BF16 vision encoder/projector. Since the official docs don't say, we can only estimate from specs: 4B parameters with Q4_K_M intuitively lands within range of typical laptops and consumer GPUs, but that's an estimate based on parameter count and quantization format, not an officially stated figure.
Buzz this issue: HF trending score 100, 5.2k downloads, 103 likes. On this issue's list, it's one of the few independent fine-tunes that isn't just someone else's weights requantized and repackaged.
Compared to similar work: The same author already has a more mature ccore/Qwen3.5-4B-3D-Coder; that version's model card says the training set was 4,146 Three.js scene examples, and FP16 weights plus a technical report are already public. The GLB version's training data spans 25 game asset domains and is focused on asset export. There's also one third-party conclusion worth reading carefully: a source called beemodel concluded "no" in its issue #434, on the grounds that this is a text-to-code model, not an image-to-3D model, so feeding it a reference image won't get you a model. The aforementioned haruspex's trial run log from 2026-11-01 reads: 8 out of 15 prompts produced a valid GLB on the first attempt, and 15/15 with a repair loop added. Both of these are third-party reports, and both likewise leave only an issue number with no full repo name or link. This show has not verified them.
How to get it: The HF page gives llama.cpp commands directly; the original text says llama serve -hf ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M to launch an OpenAI-compatible local server, or llama cli -hf ... to run straight in the terminal. The same page also lists ollama run hf.co/ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M and Docker Model Runner usage. One heads-up here: the commonly seen llama.cpp executable names are llama-server and llama-cli (the official commands quoted in items 2 and 3 of this issue use that form), so if you paste the model card text verbatim and the command isn't found, switch to llama-server -hf ... and try again.
Link: https://huggingface.co/ccore/Qwen3.5-4B-GLB-3D-Coder

2. ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF: not lower bits, just half the experts cut out

Source / summary: Picked up from Hugging Face by the "local model radar" (https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF ). The publisher is DASLab at IST Austria, and this version's selling point is its own compression methodology rather than yet another requantized repack.
What it is: A compressed GGUF built on DeepSeek-V4.1-Flash-1017 and targeted at coding ability, licensed MIT. The official model card says this is a version compressed for a specific capability: half the routed experts are pruned outright rather than quantized, while the remaining weights keep IQ2_XXS per-tensor quantization (2.00 bpw). Put plainly: everyone else is dropping bits, this one is cutting experts.
Where you'd use it:

(1) Teams with a big-memory workstation or high-end Mac who want an internal self-hosted coding assistant. Company code can't leave the building, but you still want a completion and refactoring assistant stronger than a 7B model. This kind of "trade pruning for a coder you can run offline" build exists precisely for that scenario.

(2) Internal document or UI audit pipelines that need image input. The README's startup example is llama-server paired with --mmproj mmproj-DeepSeek-V4.1-Flash-1017-BF16.gguf -ngl 99, meaning the official docs give you the mount path for the vision projector. If what you want is "dump a batch of internal system screenshots in and have it check them against UI guidelines or extract fields" without sending screens off-site, the official docs at least spell out the mount path.

(3) Engineers researching quantization and pruning who want a controlled comparison. The unpruned GSQ-RCO version in the same series notes on its card that the base model is 763.2B parameters, 384 routed experts per layer, 6 activated per token, 1M token context. Put the two repos side by side and you have ready-made teaching material on "what pruning half the experts does to capability," saving you the cost of running a training pass yourself.
Can you run it: No minimum hardware is stated officially, but the model card is very clear about file sizes: the base model is 763.2B parameters, of which 551.9B are in the transformer; the IQ2_XXS transformer weights come to roughly 72.12 GB (keeping 192 of 384 experts per layer), plus an Engram lookup file in IQ4_NL at roughly 110.60 GB, which is unpruned and identical to the unpruned version. Add the two files together and that's the order of magnitude of space you need to prepare. These are the officially listed file sizes, not an officially stated VRAM threshold. Separately, the GSQ-RCO card in the same series notes that llama.cpp b11870 or newer is required; too old a version simply won't load it.
Buzz this issue: HF trending score 97, 28.7k downloads, 101 likes. The DeepSeek-V4.1-Flash-1017 family took over half the board this issue, but most of those are pure quantization repacks. This one is among the few with a methodology of its own.
Compared to similar work: For the conventional GGUF route on the same base model, compare mrfakename's version, whose card lists Q8_0 at roughly 473 GiB, Q4_K_M at roughly 414 GiB, and even the smallest usable Q2_K at 246 GiB. DASLab prunes experts instead of merely dropping bits, and that's where the file size gap comes from. There's also an AtomicChat derivative claiming that, paired with Strata, with the backbone in system memory and the Engram table streamed from SSD, it can run on a single 12 GB GPU, but that's that repo author's own claim, and its quantization and shared Engram file came from DASLab in the first place.
How to get it: The README's flow is to fetch files with hf download ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF, link the shared Engram file as the 6th shard via ln -s, then launch with llama-cli -m IQ2_XXS/...-00001-of-00006.gguf -ngl 99 or llama-server. That symlink step is the key part; miss it and the shards won't be complete.
Link: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF

3. inclusionAI/Ling-3.1-flash-1019: a hybrid-linear MoE with 560B total parameters and only 25B activated per token

Source / summary: Picked up from Hugging Face by the "local model radar" (https://huggingface.co/inclusionAI/Ling-3.1-flash-1019 ). One of the few general-purpose large models this issue that is a first-party release with a new architecture, MIT licensed, weights directly downloadable, not somebody's repack.
What it is: inclusionAI's next-generation text-only large model. The architecture table on the official model card says Hybrid-linear MoE, 560B total parameters / 25B activated, 63 KDA layers + 9 Gated MLA layers, 512 routed experts. A note on the parameter count convention: the model card's architecture table says 560B, while the HF page sidebar shows 572B params, and the model card doesn't explain where the difference comes from, so when you cite this externally you should note which number you're using.
Where you'd use it:

(1) Teams building tool-calling agents who are picking an open-source base. The official model card says this round of post-training strengthened scientific and real-world agentic ability, listing AA-Briefcase going from 1397 to 1524 and AutomationBench-AA from 62% to 68%. If you're comparing candidates for a self-hosted agent backend, this is the official starting point for evaluation (figures are self-reported).

(2) Internal systems that need long-document or long-workflow analysis. Vercel AI Gateway's docs page states 560B total parameters, 25B activated per token, and 1M token context supporting long-document analysis and long-horizon agent work. To be concrete: a compliance department's full set of regulations plus attachments, or whole-repo Q&A across an entire codebase. If your documents routinely get their meaning mangled by chunking, 1M context at least gives you the option of "don't chunk." This show has not verified the actual quality.

(3) People who want a free trial run before deciding whether to self-host. Vercel AI Gateway lists it at free pricing, and Ling 3.1 Flash in the same series has $0.00 free routes at several providers too. Validate task feasibility via API first, then circle back to evaluate whether to pull the MIT weights on-prem. That order burns less money.
Can you run it: The official model card states no hardware threshold, only the architecture table. Check the provenance of the following two numbers carefully, one is a file size from an official FP8 release, the other is a third-party estimate for the previous generation, and neither is an official hardware requirement for the 1019 version: EXL3 quantizer vcruz305 lists inclusionAI's official FP8 release at roughly 534.35 GiB on their card; benchlm.ai estimates roughly 640GB VRAM and 8× H100 for the previous Ling 3.1 Flash. Bottom line, consumer hardware is basically out of the question, so it falls into this show's "know it exists, know what the architecture looks like" bucket.
Buzz this issue: HF trending score 90, 2.3k downloads, 94 likes. Downloads are clearly lower than the first two items, which makes sense given how few people can actually self-host it.
Compared to similar work: Compared to the DeepSeek-V4.1-Flash-1017 that's flooding this issue's board, unsloth's card describes that one as a minor update to DeepSeek-V4.1-Flash with gains on reasoning, coding, agentic, and vision benchmarks plus vision capability; Ling-3.1-flash-1019 is text-only and takes the hybrid linear attention route. inclusionAI said on X that it moved from 41 to 44 on the Artificial Analysis Intelligence Index, which is a self-reported figure.
How to get it: Downloadable safetensors weights are provided directly on HF; the page lists 572B params, tensor types BF16 / F8_E4M3 / F32, and carries the bailing_hybrid and custom_code tags, meaning loading requires trusting code inside the repo. benchlm.ai additionally notes that the previous version in the series released public FP8 weights under MIT with SGLang and vLLM serving instructions.
Link: https://huggingface.co/inclusionAI/Ling-3.1-flash-1019

Closing

Today's three items happen to stretch the late-2026 local ecosystem across its full spectrum: at one end, a 4B, Apache-2.0, narrow-task specialist you can download and play with on a laptop; at the other, a 763.2B with half its experts pruned, plus a 560B hybrid-linear MoE, the kind of beasts where it's either "you have a workstation" or "use the free API first." If you only want to tinker with one this weekend, I'd pick ccore/Qwen3.5-4B-GLB-3D-Coder: lowest barrier, clearest purpose, and it demonstrates a design philosophy more people should be copying, namely "a small model that does exactly one thing well." As always, every capability and performance claim in this episode is quoted from an official model card or a third-party report; this show has not actually run these models. See you tomorrow.