Sage-A1.5 35B (MoE)
Sage-A1.5 35B (MoE) needs 22.4 GB of VRAM at Q4_K_M with 8k context, computed from its config.json as checked on 3 November 2026. The smallest card that holds it has 24 GB, a tight fit, and the fastest desktop card that runs it, the RTX 5090, generates at most 721 tokens per second.
- Parameters
- 36B, 3B active
- Context
- 262,144 tokens
- Licence
- Apache 2.0
- Released
- October 2026
- Config checked
- 3 November 2026
Open Sage-A1.5 35B (MoE) in the LLM checker
Which card size runs it
22.4 GB against each memory size. A model may use 95% of a card, and only what is left over, or missing, is coloured; tight means above 85%, with little room for a longer conversation.
Other quantisations, context lengths or a Q8 KV cache: the VRAM calculator.
Computed 3 November 2026Which GPU to pick
The smallest card here that loads it at Q4_K_M with 8k of context is the RTX 5090 Laptop GPU (24 GB): 22.4 GB needed, an estimated 360 tokens per second, and a tight fit that leaves little for the conversation. For room to work, meaning 32,768 tokens of context while staying under 85% of memory, step up to the RTX 5090 (32 GB), which needs 22.9 GB for that and generates at about 721 tokens per second.
At Q8_0, which is near-lossless, the card is the RTX A6000 (48 GB), needing 40.4 GB. On Apple silicon, a Mac mini M4 Pro holds it with 32,768 tokens of context at an estimated 110 tokens per second; prompt processing is slower than on Nvidia, which long documents make obvious. Among unified-memory desktops, the Ryzen AI Max+ 395 holds it at 32,768 tokens in 22.9 GB, at an estimated 103 tokens per second.
Speed on each card
| GPU | Fits | Est. tokens/s | Feels like | Memory | Bandwidth |
|---|---|---|---|---|---|
| H100 SXM | Yes | ≤ 1,348 | faster than you can read | 80 GB | 3,350 GB/s |
| A100 | Yes | ≤ 820 | faster than you can read | 80 GB | 2,039 GB/s |
| RTX 5090 | Yes | ≤ 721 | faster than you can read | 32 GB | 1,792 GB/s |
| RTX PRO 6000 Blackwell | Yes | ≤ 721 | faster than you can read | 96 GB | 1,792 GB/s |
| RTX PRO 5000 Blackwell | Yes | ≤ 541 | faster than you can read | 48 GB | 1,344 GB/s |
| RTX 4090 | Tight | ≤ 406 | faster than you can read | 24 GB | 1,008 GB/s |
| RX 7900 XTX | Tight | ≤ 386 | faster than you can read | 24 GB | 960 GB/s |
| RTX 6000 Ada | Yes | ≤ 386 | faster than you can read | 48 GB | 960 GB/s |
| RTX 3090 | Tight | ≤ 377 | faster than you can read | 24 GB | 936 GB/s |
| RTX 5090 Laptop GPU | Tight | ≤ 360 | faster than you can read | 24 GB | 896 GB/s |
| RTX PRO 4500 Blackwell | Yes | ≤ 360 | faster than you can read | 32 GB | 896 GB/s |
| L40S | Yes | ≤ 348 | faster than you can read | 48 GB | 864 GB/s |
| RTX A6000 | Yes | ≤ 309 | faster than you can read | 48 GB | 768 GB/s |
| RTX PRO 4000 Blackwell | Tight | ≤ 270 | faster than you can read | 24 GB | 672 GB/s |
| Radeon AI PRO R9700 | Yes | ≤ 257 | faster than you can read | 32 GB | 640 GB/s |
| Arc Pro B70 | Yes | ≤ 245 | faster than you can read | 32 GB | 608 GB/s |
| Arc Pro B60 | Tight | ≤ 183 | faster than you can read | 24 GB | 456 GB/s |
| RTX PRO 4000 Blackwell SFF | Tight | ≤ 174 | faster than you can read | 24 GB | 432 GB/s |
| RTX Spark | Yes | ≤ 121 | faster than you can read | 128 GB (102 for the GPU) | 300 GB/s |
| DGX Spark | Yes | ≤ 110 | faster than you can read | 128 GB (126 for the GPU) | 273 GB/s |
| Ryzen AI Max+ 395 | Yes | ≤ 103 | faster than you can read | 128 GB (96 for the GPU) | 256 GB/s |
| RX 7900 XT | No | - | does not fit | 20 GB | 800 GB/s |
| RTX 4060 Ti | No | - | does not fit | 16 GB | 288 GB/s |
| RTX 5060 Ti | No | - | does not fit | 16 GB | 448 GB/s |
| RTX 5080 | No | - | does not fit | 16 GB | 960 GB/s |
| RTX 5070 Ti | No | - | does not fit | 16 GB | 896 GB/s |
| RTX 4080 SUPER | No | - | does not fit | 16 GB | 736 GB/s |
| RX 9070 XT | No | - | does not fit | 16 GB | 640 GB/s |
| RX 9060 XT | No | - | does not fit | 16 GB | 320 GB/s |
| RTX 3060 | No | - | does not fit | 12 GB | 360 GB/s |
| RTX 4070 | No | - | does not fit | 12 GB | 504 GB/s |
| RTX 5070 | No | - | does not fit | 12 GB | 672 GB/s |
| Arc B580 | No | - | does not fit | 12 GB | 456 GB/s |
Single-stream decode, from memory bandwidth at 70% efficiency. Prompt processing and batching not included. These are ceilings, not measurements: Sage-A1.5 35B (MoE) reads only 1.7 GB of weights per token at Q4, so little that the per-token costs the formula leaves out decide much of the real speed. The 21 figures marked ≤ are where a real runtime falls furthest below the number shown. See the speed estimator for other quantisations and Apple hardware, or every card compared if you are choosing hardware rather than a model.
Computed 3 November 2026Memory by quantisation and context
| Quant | 4,096 ctx | 8,192 ctx | 32,768 ctx |
|---|---|---|---|
| Q8_0 | 40.3 GB · fits 48 GB | 40.4 GB · fits 48 GB | 40.9 GB · fits 48 GB |
| Q5_K_M | 27.2 GB · fits 32 GB | 27.3 GB · fits 32 GB | 27.8 GB · fits 32 GB |
| Q4_K_M | 22.4 GB · fits 24 GB | 22.4 GB · fits 24 GB | 22.9 GB · fits 32 GB |
Weights at this quant: 20.9 GB at Q4. Every extra 1,000 tokens of context adds about 0.02 GB of KV cache at FP16, on top of a fixed 0.06 GB of recurrent state that does not grow. Try other settings in the VRAM calculator.
About the model
Sage-A1.5 35B (MoE) has 36B parameters with 3B active per token (mixture of experts). It has 40 layers, of which only 10 are full attention (2 KV heads of dimension 256) and keep a cache that grows with the conversation; the other 30 are Gated DeltaNet linear-attention layers with a small fixed state. InternScience's agent model from October 2026, built on Qwen3.6 35B-A3B and its own Agents-A1 for long-horizon scientific problem solving. Two opt-in modes set it apart: a verification mode that reviews its own progress inside its chain of thought, and orchestration of parallel sub-agents that your harness runs.
Text and images 262,144-token window Apache 2.0 October 2026 3B of 36B active
Reasoning. Reasons before it answers, returned separately from the reply. The opt-in verification mode adds a self-review to that reasoning.
Built on Qwen3.6 35B-A3B with the same expert routing: all 256 experts stay in memory and about 3B are active per token. The parameter count includes the vision encoder; a text-only GGUF is slightly smaller.
This is a reasoning model: it writes out its thinking before it answers, often thousands of tokens of it, and that thinking occupies the context window like anything else. Budget context for the reasoning, not only for the conversation.
Native context window: 262,144 tokens. Licence: Apache 2.0, a permissive licence. Weights published in October 2026.
Good at
- Long research and analysis tasks with tools, which is what it was built for.
- Parallel sub-agents: register its spawn_subagent tool and your harness runs the sub-agents it plans.
- The same memory as Qwen3.6 35B-A3B on every card, since the architecture and the expert routing are unchanged.
- Apache 2.0, with official GGUF builds from InternScience.
Watch out for
- It is a specialist built on a general model. For everyday work, Qwen3.6 35B-A3B, the model it was built on, is the plainer choice.
- Verification mode needs a non-empty system message and a JSON true, not the string "true"; without them the server returns an error.
- The 512K window is a YaRN extension you configure yourself. The native window is 262,144 tokens, and that is the figure used here.
- The card's vLLM instructions install from the main branch, not a release. The hybrid layers need a current runtime.
Similar models
The models that need about the same memory as Sage-A1.5 35B (MoE) at Q4_K_M and 8k, which makes them the real alternatives on whatever card you have. The context column is what each extra 1,000 tokens costs, and it is where models of the same size differ most.
| Model | Size | Needs | Per 1k context | Window | Released |
|---|---|---|---|---|---|
| Sage-A1.5 35B (MoE) | 36B · 3B active | 22.4 GB | 0.02 GB | 256k | 2026-10 |
| Qwen3.6 35B-A3B (MoE) | 36B · 3B active | 22.4 GB | 0.02 GB | 256k | 2026-04 |
| Qwen3 32B | 32.8B | 22.4 GB | 0.262 GB | 40k | 2025-04 |
| DeepSeek-R1 Distill Qwen 32B | 32.8B | 22.4 GB | 0.262 GB | 128k | 2025-01 |
| Qwen2.5 Coder 32B | 32.8B | 22.4 GB | 0.262 GB | 32k | 2024-11 |
See every model compared, which models fit on 8 to 96 GB, or the cheapest hour of every cloud card.
Questions
- Can I run Sage-A1.5 35B (MoE) on a 24 GB card?
- Yes, at Q4_K_M and 8k context it needs about 22.4 GB, which fits a 24 GB card tightly.
- How much VRAM does Sage-A1.5 35B (MoE) need at Q8?
- About 40.4 GB at 8k context, or 40.9 GB at 32k. Q8 is near-lossless; use it when it fits.
- What is the context length of Sage-A1.5 35B (MoE)?
- 262,144 tokens natively. Holding all of it at Q4_K_M takes about 27.6 GB, of which 5.4 GB is context.
- Why does Sage-A1.5 35B (MoE) need so little memory for long context?
- Because most of its layers do not keep a cache that grows. It has 40 layers, of which only 10 are full attention (2 KV heads of dimension 256) and keep a cache that grows with the conversation; the other 30 are Gated DeltaNet linear-attention layers with a small fixed state. At 32k tokens the context costs about 0.7 GB; if all 40 layers kept a conventional full-length cache, the same conversation would cost about 2.7 GB.
- Sage-A1.5 or Qwen3.6 35B-A3B?
- They need the same memory on every card here: Sage-A1.5 keeps the architecture and expert routing of Qwen3.6 35B-A3B. Sage-A1.5 is tuned for long scientific and research tasks, with a verification mode and sub-agent orchestration; Qwen3.6 35B-A3B is the general model it was built on, and the one for everyday work.
- How fast is Sage-A1.5 35B (MoE) on an RTX 4090?
- At most about 406 tokens per second at Q4, single stream, which is faster than you can read. That is a ceiling worked out from the card's memory bandwidth, not a measurement, and at a figure this high real runtimes land well below it, because the costs the formula leaves out take a growing share of each token.
- Architecture values from config.json in InternScience/Sage-A1.5 on Hugging Face, and its model card, checked 3 November 2026. Every figure on this page is computed from those values; check them before buying hardware for this model.
- Estimates assume a single conversation on a card that is otherwise free. A desktop on the same GPU takes 0.5 to 2 GB.