TeichAI · text · mixture of experts

GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-Distill

TeichAI/GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-Distill

GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-Distill at Q4_K_M is exactly 18,132,721,856 bytes (16.89 GiB / 18.13 GB) — an effective 4.646 bits per weight, not the nominal 4. Its KV cache at 32K is 1.65 GiB.

From the file· summed from 1 file(s)From the file· KV per layer
Parameters
31.2B
total, not active
Architecture
deepseek2
47 layers
Context
202,752
native (config.json)
License
apache-2.0

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
IQ2_M9.22 GiB9,898,321,5042.536—TeichAI
IQ3_XS11.50 GiB12,344,307,2963.163—TeichAI
Q3_K_S12.14 GiB13,034,400,9603.340—TeichAI
IQ3_M12.30 GiB13,211,266,6563.385—TeichAI
Q3_K_M13.39 GiB14,380,018,8803.685—TeichAI
IQ4_XS14.93 GiB16,033,288,8004.108—TeichAI
IQ4_NL15.79 GiB16,956,023,3924.345—TeichAI
Q4_K_M16.89 GiB18,132,721,8564.646—TeichAI
Q5_K_M19.80 GiB21,264,515,2645.449—TeichAI
Q6_K22.92 GiB24,614,786,2406.307—TeichAI
Q8_029.66 GiB31,842,799,8088.159—TeichAI
BF1655.79 GiB59,908,837,56815.351—TeichAI
F1655.79 GiB59,908,837,56815.351—TeichAI

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0960.21 GiB3.67 GiB17.78×47 / 0 / 0
8,1920.41 GiB7.34 GiB17.78×47 / 0 / 0
16,3840.83 GiB14.69 GiB17.78×47 / 0 / 0
32,7681.65 GiB29.38 GiB17.78×47 / 0 / 0
65,5363.30 GiB58.75 GiB17.78×47 / 0 / 0
131,0726.61 GiB117.50 GiB17.78×47 / 0 / 0

This model uses multi-head latent attention. No V cache is allocated at all, and the K cache stores a 512-wide latent plus 64 rope dimensions — so reading num_key_value_heads from config.json and multiplying, as every calculator does, overstates the cache by well over an order of magnitude.

Compare with

same modality, comparable size
vs Qwen3-Coder-30B-A3B-Instructvs Qwen3.6-27Bvs Qwen3.8-27Bvs gemma-4-12B-it-qat-q4_0-unquantizedvs gemma-4-E4B-itvs Qwen3-30B-A3B-Thinking-2507All alternatives →Will it fit on my card? →

Will it run on your card?

full quant x context sweep
By VRAM:8GB12GB16GB24GB32GB48GB
Radeon RX 6500 XT 4GBGeForce RTX 3050 6GBGeForce RTX 5050 8GBGeForce RTX 3080 10GBGeForce RTX 2080 Ti 11GBGeForce RTX 5070 12GBGeForce RTX 5060 Ti 16GBApple M3 Pro 18GBGeForce RTX 3080 Ti 20GBGeForce RTX 5090 D V2 24GBGeForce RTX 5090 D 32GBApple M5 Max 36GBApple M5 Pro 48GBApple M5 Max 64GBApple M3 Ultra 96GBApple M5 Max 128GBApple M2 Ultra 192GBApple M3 Ultra 256GBApple M3 Ultra 512GB

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 16.36 GiB. The real file is 16.89 GiB, because a quantization is a mixture and some tensors are always kept at higher precision. The larger discrepancy is the cache: this model allocates no value cache at all, so any formula reading num_key_value_heads overstates it by more than an order of magnitude.

Architecture

from config.json
Layers
47
Attention heads
20
KV heads
20
Head dim
256
Hidden size
2048
Vocab
154,880
Sliding window
none
SWA period
—
MLA
yes
Experts
64
Experts per token
4
use_sliding_window
—

Questions people ask

How much VRAM does GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-Distill need?
Q4_K_M is exactly 18,132,721,856 bytes (16.89 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-Distill's KV cache?
1.65 GiB at 32K context with an f16 cache, computed per layer. Quantizing the cache to q8_0 roughly halves it, which is often the difference between a context length fitting and not.
Is GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-Distill a mixture-of-experts model?
Yes — 64 experts, 4 routed per token. Every expert must be resident, but only the routed ones are read per token, which is why its memory requirement and its speed behave very differently.
Which quantization of GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-Distill should I use?
Q4_K_M is the usual default. Pick the largest quantization that fits your card at the context you actually need — the table above gives exact sizes for every one published.