GLM-5.3-Flash-colibri-int4-g64

GLM-5.3-Flash for colibrì (int4-gs64)

An int4 group-scaled conversion of zai-org/GLM-5.3-Flash, packed in the container format that colibrì streams from disk.

This is not the official checkpoint. The weights are Z.ai's; this repository only changes how they are stored. It will not load in transformers or vLLM. It loads in colibrì's glm53 engine and nothing else.

Text and vision both work.

What it is for

Running a 321B multimodal model on a machine with 25 GB of RAM, by keeping the routed experts on disk and reading only the ones each token selects.

Those decode numbers are honest and they are the point of the next section.

Read this before you download 195 GB

The disk is the wall. One token touches 42 sparse layers times 8 experts times 14.2 MB, which is 4.8 GB of weights read per token. Measured with O_DIRECT, the reference drive saturates near 200 MB/s, so there is a floor of roughly 24 seconds per token on that hardware. No CPU and no GPU moves it. Faster storage does; fewer bytes would.

So this is a model you can run, not one you can chat with quickly. If that trade is not what you want, the smaller families colibrì supports (OLMoE, Qwen3.6) are interactive on ordinary hardware.

Use

git clone https://github.com/JustVugg/colibri && cd colibri/c
make glm53

./coli chat  --model /path/to/this/repo --no-think
./coli serve --model /path/to/this/repo
./coli web   --model /path/to/this/repo

Images

Paste a path straight into coli chat, attach or drop a file in coli web, or send an OpenAI image_url part with a base64 data URI or a local path.

GLM53_MAX_IMAGE_TOKENS is the setting that matters. The checkpoint's own ceiling is 8000 tokens per image, which on a disk-streaming engine is a prefill nobody will sit through. 256 keeps ordinary text legible, 64 keeps shapes and colours. The image is shrunk, not cropped.

Reasoning

The generation prompt opens <think> and the model closes it, which is what the official template does. --no-think closes it immediately and the answer starts at the first word. --effort low|high|max picks the level the template understands. Both are time controls on this engine rather than matters of taste.

What the conversion did

Routed experts become int4 group-scaled at 64: name as U8 packed nibbles plus name.qs as F32 scales, stored flat, with shapes taken from config.json. Everything else stays BF16, which is 9.7B parameters of 321 and about 18 GB, so the precision of the dense set is a load-time choice and retuning it never means downloading the source repository again.

Every tensor kind is classified explicitly during conversion and an unrecognised name stops it, because a converter that silently skips what it does not know produces a checkpoint that loads and is quietly missing a tensor.

Reproduce it with:

python3 tools/convert_glm53.py --outdir /path/glm53_i4 --min-free-gb 30

On the reference machine that was 62 shards, 194.7 GB out, about 25 hours. Peak disk is the output plus a single 5 GB source shard, never the source repository's 328 GB.

Architecture, as read from this container

45 text layers plus one MTP, hidden 4096, vocab 154880, context up to 1,048,576. 34 layers use KDA linear attention and 11 use DSA. Attention is MLA with q_lora_rank 1536, kv_lora_rank 512, 64 heads, qk_nope_head_dim 256 and qk_rope_head_dim 0, so there is no rotary component. 288 routed experts with top-8 selection plus one shared expert, moe_intermediate_size 2048, the first 3 layers dense. A 24-block ViT tower at patch 14 and 448 px.

The KV cache is absorbed rather than expanded. kv_b_proj is folded into both ends, so the cache holds the 512-wide latent instead of expanded keys:

score_j = q · (W_k c_j) = (W_kᵀ q) · c_j
out     = Σ_j a_j (W_v c_j) = W_v (Σ_j a_j c_j)

An identity, not an approximation, and the reason the KV state is 33 KB per token instead of 1.39 MB.

Quality

No independent benchmark of this container has been published yet. On GLM-5.2, group-scaled int4 measured within noise of int8 on packaged benchmarks, while per-row int4 scales cost materially more, which is why this uses groups of 64. If you measure this one, please open an issue on the colibrì repository with the numbers.

Licence and attribution

MIT, inherited from zai-org/GLM-5.3-Flash. The model is Z.ai's work; this repository redistributes a requantised copy of their weights and claims nothing beyond the conversion.

Engine: github.com/JustVugg/colibri · engine notes: docs/glm53-flash.md