# DeepSeek R1 0528 NVFP4 vs Qwen3 235B A22B Instruct 2507 FP8 Throughput

On provider list prices, Qwen3 235B A22B Instruct 2507 FP8 Throughput costs $0.20 per million input tokens against $3 for DeepSeek R1 0528 NVFP4: 15.0x apart. Output is $0.60 against $7 (11.7x). On Allocate both bill at list plus the 7% transaction fee.

## Specifications

| | DeepSeek R1 0528 NVFP4 | Qwen3 235B A22B Instruct 2507 FP8 Throughput |
| --- | --- | --- |
| Lab | Deepseek | Qwen |
| Access | Open weights | Open weights |
| Context window | 160K tokens | 256K tokens |
| List price, input | $3 / M tokens | $0.20 / M tokens |
| List price, output | $7 / M tokens | $0.60 / M tokens |
| Cached input | n/a | n/a |
| License | MIT | Apache 2.0 |
| Fine-tunable | Yes | Yes |

Specifications and provider list prices from the Allocate catalog, checked 2026-07-08. Billed price is list plus the 7% transaction fee.

## What the numbers say

Take 1,000,000 requests a month at 1,200 input and 350 output tokens each. That workload costs $450 a month on Qwen3 235B A22B Instruct 2507 FP8 Throughput and $6,050 on DeepSeek R1 0528 NVFP4 at list: a gap of $5,600, or 13.4x.

Qwen3 235B A22B Instruct 2507 FP8 Throughput reads 256K tokens per request against 160K for DeepSeek R1 0528 NVFP4, 1.6x the window. That decides which one can take whole documents without splitting them.

## Choose DeepSeek R1 0528 NVFP4 for

- Fine-tuning under a permissive license (MIT)

## Choose Qwen3 235B A22B Instruct 2507 FP8 Throughput for

- The lower list price ($0.20 in / $0.60 out per M tokens)
- The longer context window (256K vs 160K tokens)
- Fine-tuning under a permissive license (Apache 2.0)

## Common questions

### Which is cheaper, DeepSeek R1 0528 NVFP4 or Qwen3 235B A22B Instruct 2507 FP8 Throughput?

Qwen3 235B A22B Instruct 2507 FP8 Throughput, on this workload shape. At list prices it is $0.20/$0.60 per million tokens in and out against $3/$7 for DeepSeek R1 0528 NVFP4. Billed on Allocate: $0.21/$0.64 against $3.21/$7.49, list plus 7%.

### Which has the bigger context window?

Qwen3 235B A22B Instruct 2507 FP8 Throughput: 262,144 tokens (256K) against 163,840 (160K) for DeepSeek R1 0528 NVFP4.

### Can I fine-tune DeepSeek R1 0528 NVFP4 or Qwen3 235B A22B Instruct 2507 FP8 Throughput?

Both publish open weights (DeepSeek R1 0528 NVFP4: MIT; Qwen3 235B A22B Instruct 2507 FP8 Throughput: Apache 2.0), so both can be fine-tuned. On Allocate the trained weights stay inside your boundary and belong to you.

---

[HTML page](https://allocate.network/compare/deepseek-deepseek-r1-0528-vs-qwen-qwen3-235b-a22b-instruct-2507-tput) · [DeepSeek R1 0528 NVFP4](https://allocate.network/models/deepseek-deepseek-r1-0528.md) · [Qwen3 235B A22B Instruct 2507 FP8 Throughput](https://allocate.network/models/qwen-qwen3-235b-a22b-instruct-2507-tput.md) · [Machine-readable catalog](https://allocate.network/catalog.json)