Hardware Implementation of Qwen3.5 on FPGA Fabric

A developer has reported successful implementation of the Qwen3.5 large language model (LLM) architecture on Field Programmable Gate Array (FPGA) hardware, utilizing relatively low-cost components. According to a report shared on Reddit, the implementation focuses on the 9B and 27B parameter versions of the model using INT4 quantization.

The project utilizes SQRL FK33 boards, which feature 8GB of HBM2 memory with approximately 400GB/s bandwidth, and “Jungle Cat” ex-mining FPGAs. In unverified testing results, the developer reported that the Qwen3.5-9B INT4 model achieved a generation speed of approximately 2.4 to 3.2 tokens per second (tok/s) at 75MHz, depending on the context length.

The developer also provided theoretical projections for the Qwen3.5-27B INT4 model. Using a configuration of four XCVU35P dies with 4-way tensor parallelism at 200 MHz, the model is estimated to reach 25 tok/s for both prefill and generation at short context, though performance is expected to drop to approximately 1 tok/s at a 262k context window due to KV cache limitations.

The implementation is currently hosted in an open-source repository on GitHub under the MIT license. The developer noted that further optimization, such as increasing clock speeds or moving to an ASIC (Application-Specific Integrated Circuit) on a TSMC 2023 process node, could significantly increase throughput.

Developments in the Agentic Economy and Solana Infrastructure

In the broader landscape of autonomous machine commerce, new infrastructure is emerging to support “agentic” transactions. Reports from the Solana ecosystem indicate a shift toward micro-latency optimization for autonomous agents.

Key infrastructure developments include the evolution of Pay.sh, which provides full-stack infrastructure for machine commerce, allowing agents to access paid APIs via the x402 standard. Additionally, PayAINetwork has implemented x402 batch settlement on the Solana blockchain, a move intended to allow merchants to consolidate thousands of individual micropayments into single claims.

On the corporate and consumer side, several integrations have been reported:

  • Moonpay and PayBox have launched a desktop application for Mac and integrated with the Cursor AI marketplace, providing execution rails for agents such as Claude and ChatGPT.
  • Coinbase Developer has upgraded its Solana facilitator, reporting a 66% increase in verification speed and a 20% improvement in settlement speed.
  • Laso Finance has opened card on-ramps for both humans and agents through Crossmint.

Specialized AI Tooling and Market Observability

The ecosystem is seeing an increase in specialized tools designed for autonomous workflows. BlockRunAI has released OpenJev, a 4B cross-encoder designed to return specific labels or scores rather than generative text.

To manage the growing volume of machine-to-machine transactions, new observability tools have emerged. Agentrateindicators.com has launched to track real-time API rate metrics and settlement telemetry. Furthermore, Dexteronsol’s Indexter has mapped over 170 active x402 endpoints to assist autonomous callers in discovering service providers.

In the application layer, several projects are targeting specific niches. SteveTheAgentAI is reportedly managing over 170 agents and 2,900 transactions on-chain, while SelfMadebySP3ND is an agent designed to generate revenue specifically to fund its own hardware requirements.