Skip to main content
Back

George Hotz's tiny corp Launches Qwen API at 47% Below Alibaba's Price

George Hotz's tiny corp is launching a hosted Qwen 3.8 Flash API at half the official price, served from AMD 7900XTX tinyboxes running tinygrad.

Repo, Full Access
  • The tiny corp is launching a hosted inference API for Qwen 3.8 Flash at half the official price.
  • Output tokens will cost roughly 25 cents per million, versus $0.47 on Alibaba's API.
  • New users get $25 in launch credits, plus $1 free for signing up via GitHub.
  • Vision support is being added before launch, enabling screenshot-driven agent workloads.
  • Backend runs on 6x AMD 7900XTX tinyboxes using tinygrad's new serve.py stack.
  • George Hotz frames it as a learning exercise, hinting Qwen 4 Flash may ship on day one.

tiny corp, George Hotz’s company behind the tinygrad framework and AMD-based tinybox systems, plans to launch a hosted API for Qwen model files on the Friday following its X announcement. The company quoted $0.25 per million output tokens and said it may substitute Qwen4-Flash if that model ships in time. Because the post gave a relative launch date, developers should verify current availability and the deployed model before integrating.

The service gives tiny corp a live environment for testing its inference stack on consumer AMD GPUs. The company described the launch as an operational experiment, with pricing intended to limit abuse while attracting enough traffic to expose performance and reliability problems.

A quarter per million output tokens

A pricing reference lists Alibaba’s hosted Qwen3.8-Flash endpoint at $0.15 per million input tokens and $0.47 per million output tokens. tiny corp’s quoted output rate is about 47% lower.

Usage Alibaba-hosted rate tiny corp rate
Input tokens $0.15 per million About $0.075 if the half-price claim applies
Output tokens $0.47 per million $0.25 per million
Cached input About 89% below the standard input rate Undisclosed

The announcement explicitly specified the output price. Input pricing, cache discounts, minimum charges, and billing granularity still require confirmation. At the quoted output rate, $1 buys up to four million output-only tokens, while $25 buys up to 100 million.

Free credits, vision, and few guarantees

The initial offer combines developer credits with a limited set of product commitments:

  • Builder credit: $25 at launch for developers building against the API. The post cited the PokéBench harness as an early testing workload.
  • Signup credit: $1 for each user who registers with a GitHub account. The announcement did not specify whether this credit stacks with the builder allocation.
  • Vision support: Multimodal requests are planned before launch, enabling screenshot and image-based agents.
  • Latency target: The company expects performance to resemble Alibaba’s endpoint under load and improve when spare capacity is available.

The announcement omitted rate limits, concurrency caps, context length, supported regions, data-retention terms, uptime targets, and support arrangements. It also left OpenAI API compatibility details unclear, including streaming, tool calls, structured output, token counting, and error behavior.

A shared demonstration showed Qwen3.8-Flash-Next guiding a Pokémon agent through Viridian Forest. That workload exercises visual state interpretation, action selection, and repeated tool calls, though a single run provides no latency distribution, completion rate, or reproducibility data.

Six Radeon cards behind the endpoint

A tinygrad pull request titled Qwen 3.8 Next Flash on 6x7900XTX adds substantial serving code to tinygrad/llm/serve.py and a new engine.py. The changes point to a configuration using six Radeon RX 7900 XTX cards, with tinygrad compiling and scheduling the workload through its own stack while bypassing AMD’s HIP and ROCm user-space compute layers.

The approach follows tiny corp’s hardware pitch. Its published comparison describes a ten-card 7900 XTX system with 1,230 BF16 TFLOPS, 240GB of VRAM, and 9,600GB/s of aggregate memory bandwidth for about $9,990. Those peak figures provide a hardware-cost comparison; sustained token throughput also depends on quantization, memory movement, utilization, power, cooling, networking, and batching.

Sparse compute, substantial memory

Qwen3.8-Flash-Next is an experimental multimodal mixture-of-experts model with 125 billion total parameters, about six billion active parameters per token, and a 51B n-gram table for common token sequences. Qwen says the architecture previews work planned for the Qwen4 family, giving inference engines and tool vendors time to add support. The weights use the Qwen Community License, which teams should review before commercial deployment or redistribution.

Sparse activation reduces the computation required for each generated token, but the full expert set still needs storage or an efficient loading strategy. Six 7900 XTX cards provide 144GB of aggregate VRAM, while 125 billion parameters would require roughly 250GB for 16-bit weights before runtime overhead. The proposed configuration therefore depends on quantization, sharding, and careful memory management.

Artificial Analysis ranked the model fifth among 111 models in its cited class, while also recording unusually long responses. Its Intelligence Index run consumed about 200 million output tokens, compared with a 110 million median. At $0.47 per million output tokens, that volume costs about $94 on Alibaba’s endpoint; tiny corp’s quoted rate would reduce the same output charge to $50, excluding input and cached tokens. Developers should still compare total task cost because a verbose model can consume more tokens than a concise competitor.

Before routing production traffic

Developers evaluating the API should test the service against their own agent, retrieval, and multimodal workloads before choosing it as a primary provider:

  1. Pin the model: Confirm whether the endpoint serves Qwen3.8-Flash-Next or Qwen4-Flash, and determine how model changes will be announced.
  2. Verify the bill: Check input, output, cache, image, and retry charges, along with credit expiration and spending limits.
  3. Test compatibility: Exercise streaming, tool calls, JSON output, vision payloads, context limits, tokenization, and error handling.
  4. Measure performance: Record throughput, time to first token, median latency, tail latency, queueing behavior, and output length under concurrent load.
  5. Review operations: Confirm rate limits, uptime commitments, status reporting, regional availability, support channels, logging, retention, and training policies.
  6. Keep a fallback: An alternate provider or self-hosted deployment can protect production systems from capacity limits and experimental model changes.

The free credits make meaningful evaluations inexpensive, while production suitability will depend on sustained throughput, tail latency, uptime, and total task cost once external traffic reaches the service. The launch will also show whether tiny corp can translate its AMD and tinygrad architecture into a reliable hosted inference product.