DeepInfra raises $107M Series B to scale the inference cloud — read the announcement

Qwen3.5 35B A3B is a native vision-language model released by Alibaba Cloud in February 2026. It uses a hybrid architecture that integrates Gated Delta Networks with a sparse Mixture-of-Experts model, achieving higher inference efficiency. With 35 billion total parameters and only 3 billion activated per token through 256 experts (8 routed + 1 shared), it outperforms previous-generation models more than 6x its size.
The model supports a 262k token context window (extensible to 1M via YaRN), dual thinking and non-thinking modes, tool calling, and 201 languages and dialects. Qwen3.5-Flash is the hosted API version corresponding to Qwen3.5-35B-A3B, offering additional production features including 1M context length by default and official built-in tools. The model is released under the Apache 2.0 license.
Qwen3.5 35B A3B is being offered by multiple providers, but they’re not created equal. This analysis breaks down which one delivers the best performance, lowest cost, and fastest response times for your use case.
| Provider | Blended ($/1M) | Speed (t/s) | Latency (TTFT) | E2E 500 Tokens (s) |
|---|---|---|---|---|
| DeepInfra (FP8) | $0.71 | 175 | 0.60s | 14.86s |
| Novita | $0.69 | 184 | 1.73s | 15.35s |
| GMI (FP8) | $0.69 | 190 | 2.39s | 15.57s |
| Alibaba Cloud | $0.69 | 162 | 2.04s | 17.48s |
Based on benchmarks across tracked providers, DeepInfra is the recommended API for production-scale Qwen3.5 35B A3B deployment. Its unmatched TTFT of 0.60s and best-in-class end-to-end response time (14.86s) make it the top choice for interactive and user-facing applications. For workloads prioritising maximum raw throughput, GMI (FP8) leads at 190 t/s. For the most cost-sensitive deployments, Novita, GMI, and Alibaba Cloud are all tied at $0.69/1M blended.
DeepInfra stands out as the overall recommended provider, delivering the lowest latency and the fastest end-to-end response time among all evaluated options.
While its token generation pace remains highly competitive at 175 t/s, it is the exceptionally brief TTFT of 0.60s that makes DeepInfra the superior choice for interactive applications such as real-time assistants, conversational AI, and coding tools. Its blended price of $0.71/1M is marginally above the $0.69 floor, but the performance advantage more than justifies the difference for user-facing workloads.
For workloads where the primary requirement is the highest volume of tokens generated per second, GMI (FP8) leads the benchmark.
GMI’s 190 t/s throughput is the highest measured, making it the natural choice for batch processing, offline data generation, or summarization tasks where the initial latency is not a critical constraint. Its TTFT of 2.39s, however, makes it less suitable for real-time user-facing applications where perceived responsiveness matters.
Novita offers a compelling middle ground between generation speed and initial responsiveness, making it a versatile option for mixed workloads.
Novita ranks as a strong runner-up in both throughput (184 t/s, #2) and end-to-end response time (15.35s, #2). Combined with the lowest blended price in the benchmark ($0.69/1M), it is an excellent choice for developers who need a reliable middle ground between DeepInfra’s interactivity and GMI’s raw throughput, without paying a premium.
As the model creator, Alibaba Cloud offers a reliable first-party baseline with full production support via the Qwen3.5-Flash hosted API.
Alibaba Cloud’s pricing matches the market floor at $0.69/1M, but its throughput (162 t/s, lowest in the benchmark) and end-to-end response time (17.48s, slowest) make it the least performant option for pure inference speed. It remains the natural fallback for teams already in the Alibaba Cloud ecosystem or needing the extended production features of Qwen3.5-Flash.
Selecting the right provider for Qwen3.5 35B A3B comes down to your application’s primary bottleneck. For interactive, user-facing applications where every millisecond counts, DeepInfra’s unmatched TTFT (0.60s) and end-to-end performance (14.86s) make it the standout choice. For batch workloads requiring maximum sustained throughput, GMI (FP8) leads at 190 t/s. For teams needing a cost-efficient balance of both, Novita delivers strong performance across the board at the market’s lowest price.
How to deploy google/flan-ul2 - simple. (open source ChatGPT alternative)Flan-UL2 is probably the best open source model available right now for chatbots. In this post
we will show you how to get started with it very easily. Flan-UL2 is large -
20B parameters. It is fine tuned version of the UL2 model using Flan dataset.
Because this is quite a large model it is not eas...
How Mixture of Experts Models Changed LLM Economics<p>Every open-weight model that has closed the gap with GPT-5.5 and Claude Opus 4.7 this year has one thing in common. DeepSeek V4-Pro: 1.6 trillion parameters, 49 billion active per token. Kimi K2.6: 1 trillion parameters, 32 billion active. GLM-5.1: 744 billion parameters, 40 billion active. MiniMax M2.7: large total parameter count, 10 billion active […]</p>
Inference Economics: True AI Costs at Scale<p>Most teams discover their inference economics the same way: a production bill arrives that looks nothing like the number they expected. The per-token price seemed small enough during testing. Then real traffic showed up, agents started chaining calls, RAG pipelines bloated the context window, and suddenly the math looked completely different. Token prices have fallen […]</p>
© 2026 DeepInfra. All rights reserved.