Qwen3-235B: the model previously served by Gonka

Qwen3-235B-A22B-Instruct-2507-FP8 is a Large Language Model (LLM) from the Qwen3 family, developed by the Qwen team at Alibaba Cloud. The full name deciphers as: Qwen3 — the third generation of the series, 235B — 235 billion parameters in total, A22B — 22 billion active parameters per query, Instruct — a version trained to follow instructions, 2507 — July 2025 release, FP8 — 8-bit quantization for memory optimization.

A key architectural feature is MoE (Mixture of Experts). Unlike "dense" models (GPT-5.5, Claude Sonnet 4.6), where each token passes through all parameters, an MoE model activates only a subset of "experts" — specialized blocks of the neural network — for each query. In the case of Qwen3-235B, out of 235 billion parameters, only 22 billion are activated for each token — less than 10%. This delivers quality comparable to models with 200B+ parameters at the computational cost of a 22B model.

Practically, this means: the model is smarter than one might expect from its speed. It processes queries significantly faster than dense models of comparable quality, while requiring orders of magnitude less VRAM for inference. This is why MoE has become the dominant architecture for the largest models of 2025–2026.

The context window for Qwen3-235B is 131,072 tokens (~100,000 words) — sufficient for analyzing entire books, codebases, or long legal documents in a single query. The model supports 119 languages, including Russian, English, Chinese, Arabic, Hindi, and dozens of others — making it one of the most multilingual models on the market.

Qwen3-235B competes with the largest proprietary and open models. Here is a comparison of key characteristics:

Qwen3-235B demonstrates quality levels comparable to GPT-5.5 and Claude Sonnet 4.6 on most benchmarks, while being an open-weights MoE model that is exponentially cheaper than proprietary analogs: MoE architecture activates only a portion of parameters per request, drastically reducing compute costs. The same principle of cheap decentralized inference is being applied by the Gonka network today to its live models — Kimi K2.6 and MiniMax M2.7, accessible via JoinGonka Gateway at $0.003 per 1M tokens (thousands of times cheaper than GPT-5.5 and Claude).

On the MMLU-Pro, HumanEval, MATH-500, and GSM8K benchmarks, the model ranks in the top three open-source models, trailing only DeepSeek-R1 in mathematical reasoning tasks. In code generation, translation, and instruction following, Qwen3-235B consistently outperforms Llama 4 Maverick and is comparable to Claude Sonnet 4.6.

When Qwen3-235B was the network's primary model, it ran on the Gonka network in a distributed manner via the DiLoCo protocol, adapted for inference. The full model in FP8 format requires about 640 GB of video memory (VRAM), which is impossible to fit onto a single GPU — even an H100 80GB or H200 141GB is not enough. Therefore, the model was split across layers (tensor parallelism + pipeline parallelism) between several ML-nodes.

In practice, Qwen3-235B ran on a cluster of 8-16 GPU-nodes, each with at least 40 GB of VRAM. Transfer Agents routed the request to the required cluster, vLLM on each node processed its share of the model, and the results were aggregated and returned to the user. The entire process took hundreds of milliseconds — the user did not feel that their request was processed by a dozen GPUs in different parts of the planet. The network applies the same principle of distributed inference today — to its currently active models.

An important technical detail: Gonka uses vLLM as the serving engine. vLLM is an open-source project that provides high-performance text generation via PagedAttention — an algorithm that optimizes memory usage when processing multiple requests in parallel. This allows the network to serve thousands of concurrent users without quality degradation.

The model supports native tool calling — calling functions and tools directly from the model's response. This capability was added to Gonka via PR #767 with a 0.958 threshold for tool call detection. On Qwen3-235B, this allowed developers to build AI agents that interact with external APIs, databases, and tools — all through a single request. Tool calling support has been maintained in the network's current models.

The Gonka network comprises over 4,000 GPUs (H100, H200, A100, RTX 4090, and others) organized into 120+ ML-nodes. It is one of the largest distributed GPU networks for AI inference in the world — and if this power was previously directed at Qwen3-235B, today it serves the network's active models, Kimi K2.6 and MiniMax M2.7.

Direct answer: through the Gonka network — no longer. Qwen3-235B has been removed from the network, so you can no longer invoke it by the qwen3-235b-a22b identifier via our Gateway. Since the model is open (Apache 2.0), you can still run it yourself or access it via third-party open-weights model hosting services — for example, via OpenRouter (estimated $0.071/$0.100 per 1M tokens).

If you need the very same cheap decentralized inference that brings people to Gonka, it has not gone anywhere; it is just working on current network models now. Via JoinGonka API Gateway, Kimi K2.6 and MiniMax M2.7 are available through an OpenAI-compatible API: any code written for OpenAI works without changes — you just need to replace the URL, API key, and model name.

Example request to a current model:

curl https://gate.joingonka.ai/api/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moonshotai/Kimi-K2.6",
    "messages": [{"role": "user", "content": "Explain the MoE architecture"}]
  }'

Cost: $0.003 for 1 million tokens — this is 1,700 times cheaper than GPT-5.5 ($5.00/1M) and 1,000 times cheaper than Claude Sonnet 4.6 ($3.00/1M). Upon registration, you receive 10 million tokens for free for testing.

The Gateway is compatible with popular development tools: Quick Start describes connectivity via Python, Node.js, and curl. IDE integrations are also supported — Cursor, Continue, Cline, Aider, and Claude Code — as well as AI agent frameworks: LangChain, n8n, LibreChat, Open WebUI.

For a quick start:

1. Register at gate.joingonka.ai (connect a wallet or create a new one)
2. Get an API key in the Dashboard
3. Replace api.openai.com with gate.joingonka.ai/api in your code
4. Specify the current network model — moonshotai/Kimi-K2.6 or MiniMaxAI/MiniMax-M2.7 (full list in the GET /v1/models endpoint)

Decentralized enterprise-grade inference at hobby-project prices has not gone anywhere — Qwen3-235B just paved the way for newer network models. The same price-to-quality ratio now powers Kimi K2.6 and MiniMax M2.7.