Sunk Cost sunkcost.ai Data checked 2026-09-03

What hardware do you need to run Sage S2.1 Flash?

Sage S2.1 Flash at Q4_K_M is 194 GB of weights, with a context ceiling of 1M tokens. A scientific model from Shanghai Artificial Intelligence Laboratory, built for work that runs several rounds of analysis and checking. It reads images and video, and projectors published beside it feed it molecular graphs, crystal structures and protein sequences, none of which is in the file sizes here. Its config.json matches GLM-5.3-Flash's field for field, down to the 45 layers, the 288 experts with eight of them running a token and the 24-layer vision tower, so the memory arithmetic on this page is that model's. The speeds here are worked out from memory bandwidth, the way every estimate on this site is.

Nothing on the market runs it at 32kAt 32k context it needs 194 GB, more than any machine here with a published price that is still sold hands a model.
What does hold itThe two machines here sold with 512 GB, which hand a model 384 GB: one no longer sold and one with no published price.

How good is Sage S2.1 Flash, really?

It has not been placed on the intelligence index yet. See the whole table.

Head to head with a model that needs much the same memory at 32k of context: vs Tencent Hy3.

What Sage S2.1 Flash costs either way

Nobody rents Sage S2.1 Flash by the token. The closest hosted match, GLM-5.3-Flash, costs $0.0399 per million input tokens and $0.0784 per million output (OpenRouter, cheapest active endpoint, checked 2026-11-18). Buying a machine only beats that if you use it hard enough, for long enough, that the hardware price divides down below the rental bill. What a million tokens costs each way puts the two prices side by side.

The machines that miss Sage S2.1 Flash, and what they run

None of the machines here with a published price that are still sold holds Sage S2.1 Flash at any window, and none of them misses by a window: its 194 GB of weights are larger than the usable memory in every one of them, before a single token of key-value cache. The Mac Studio M5 Ultra, 256GB comes nearest, 1.9 GB short at 4k, the shortest window on the list.

Two machines here do hold Sage S2.1 Flash at 32k, and neither of them is an answer to what to buy, which is the question the rest of this page asks: the Mac Studio M3 Ultra, 512GB, no longer sold, at the $9,499 it launched at, and the Mac Studio M5 Ultra, 512GB, on sale with no published price. They are sold with 512 GB of memory and hand a model 384 GB of it, against the 194 GB Sage S2.1 Flash needs at 32k.

MachinePriceUsable memoryShort byStrongest model it holds
Mac Studio M5 Ultra, 256GB $10,799 192 GB 1.9 GB GLM-5.3-Flash
DGX Spark GB10 Grace Blackwell, 128GB $4,699 119.5 GB 74.4 GB Qwen3.8 27B
Strix Halo Framework Desktop, 128GB $3,449 96 GB 97.9 GB Qwen3.8 27B
MacBook Pro M5 Max (16-inch), 128GB $6,999 96 GB 97.9 GB Qwen3.8 27B
NVIDIA RTX PRO 6000 Blackwell, 96GB $18,000card only 95 GB 98.9 GB Qwen3.8 27B
Mac mini M5 Pro, 64GB $2,699 48 GB 145.9 GB Qwen3.8 27B
AMD Radeon AI PRO R9700, 32GB $1,299card only 31 GB 162.9 GB Qwen3.8 27B
MacBook Air M5 (13-inch), 16GB $1,299 10.5 GB 183.4 GB Gemma 4 12B

One machine per family, the largest configuration in each that still misses, nearest first. Short by is the weights plus the key-value cache at 4k measured against the memory that machine's GPU can address, which that machine's own page gives. The last column is the highest-scoring current model the machine holds at 32k.

Price the Mac Studio M5 Ultra, 256GB on GLM-5.3-Flash instead

The specifics

Parameters
321.3B, of which 18B are active per token
Quantisation
Q4_K_M
Weights on disk
194 GB
Images
It reads images as well as text, and the weights above are the language model. Feeding it pictures loads a separate 0.7 GB vision projector, Sage-S2.1-Flash.mmproj-Q8_0.gguf, on top of every figure on this page.
KV cache
0.6 GB at 32k context — 34 of 45 layers use linear attention with a fixed state; the other 11 use latent attention with no RoPE part (1,024 B) plus a sparse-attention indexer (514 B). The config states the same 45 layers, the same 512-wide latent and the same 128-wide indexer as GLM-5.3-Flash, field for field. Note this is the figure for vLLM or SGLang: the reference Transformers code expands the latent before caching and uses far more. How weights and cache add up.
Maximum context
1M tokens
Licence
MIT
Sources
internlm/Sage-S2.1-Flash, internlm/Sage-S2.1-Flash (config), mradermacher/Sage-S2.1-Flash-GGUF, zai-org/GLM-5.3-Flash, openrouter.ai