Spaces:

IteraCompute
/

README

Running

App Files Files Community

liwhattao commited on 15 days ago

Commit

f8af606

verified ·

1 Parent(s): aa6e82b

Announce Qwen3.8 NVFP4 release and forthcoming Qwen 4 NVFP4

Browse files

Files changed (1) hide show

README.md +91 -82

README.md CHANGED Viewed

@@ -1,82 +1,91 @@
----
-title: README
-emoji: 🦀
-colorFrom: purple
-colorTo: pink
-sdk: static
-pinned: false
----
-# IteraCompute
-**OpenAI-compatible inference for open-weight models. Start with free Qwen3.8-27B.**
-IteraCompute Artificial Intelligence Technology (Chongqing) Co., Ltd. provides reliable, usage-based inference for language models through OpenAI-compatible APIs. We focus on predictable integration, streaming responses, tool use, structured outputs, and privacy-conscious processing.
-## Start with free Qwen3.8-27B
-**Qwen3.8-27B is available for free through IteraCompute's OpenAI-compatible API.** Input, cached input, and output are all **$0 per 1M tokens**.
-- API model ID: `qwen/qwen3.8-27b`
-- API base URL: `https://api.iteracompute.com/v1`
-- Free tier: up to **30M tokens per account per UTC calendar month**, **128K total context**, and **160,000 TPM**.
-- Supports Chat Completions, streaming, and tool calling. Registration and email verification are required; no card or recharge is needed for the Free tier.
-The monthly allowance counts input and output, including reasoning, across the account's API keys. Free capacity is shared and may queue at busy times; fair-use limits apply and there is no uptime or latency SLA. See the current [free allowance and limits](https://iteracompute.com/free/) and [Free Tier Terms](https://iteracompute.com/free-tier-terms/).
-[Start free](https://iteracompute.com/free/) · [First API call](https://iteracompute.com/docs.html) · [Model details](https://iteracompute.com/models/qwen3-8-27b/)
-中文：Qwen3.8-27B 目前免费，输入、缓存输入和输出均为 $0；免费档每月最高 30M tokens、128K 上下文、160K TPM，无需信用卡或充值，高峰期可能排队。具体额度与可用性以官网和控制台为准。
-## Current model API catalogue
-The [website model catalogue](https://iteracompute.com/models/) lists the following models as online, checked on **2026-10-14**. These are direct IteraCompute API IDs; use them exactly as shown. Current availability, capabilities, and pricing are published on the website and in the Console.
-| Model | API model ID | Access |
-| --- | --- | --- |
-| [Qwen3.8-27B](https://iteracompute.com/models/qwen3-8-27b/) | `qwen/qwen3.8-27b` | Free within published allowance |
-| [Qwen3.8-2.4T-A95B](https://iteracompute.com/models/qwen3-8-2-4t-a95b/) | `qwen/qwen3.8-2.4t-a95b` | Usage-based |
-| [GLM-5.3-Flash](https://iteracompute.com/models/glm-5-3-flash/) | `z-ai/glm-5.3-flash` | Usage-based |
-| [Kimi K2.7 Code](https://iteracompute.com/models/kimi-k2-7-code/) | `moonshotai/kimi-k2.7-code` | Usage-based |
-| [DeepSeek V4 Pro](https://iteracompute.com/models/deepseek-v4-pro/) | `deepseek/deepseek-v4-pro-0813` | Usage-based |
-| [GLM-5.3](https://iteracompute.com/models/glm-5-3/) | `z-ai/glm-5.3` | Usage-based |
-| [Kimi K3](https://iteracompute.com/models/kimi-k3/) | `moonshotai/kimi-k3` | Usage-based |
-| [MiniMax M3](https://iteracompute.com/models/minimax-m3/) | `minimax/minimax-m3` | Usage-based |
-| [DeepSeek V4.1 Flash](https://iteracompute.com/models/deepseek-v4-1-flash/) | `deepseek/deepseek-v4.1-flash` | Usage-based |
-| [DeepSeek V4 Flash 0731](https://iteracompute.com/models/deepseek-v4-flash-0731/) | `deepseek/deepseek-v4-flash-0731` | Legacy alias for V4.1 Flash |
-[Current pricing](https://iteracompute.com/pricing/) · [Public Models endpoint](https://api.iteracompute.com/v1/models)
-## Open-source quantization
-### Qwen4-27B NVFP4 — preparation stage
-We are preparing a reproducible NVFP4 quantization and release workflow for the anticipated Qwen4-27B model. The toolkit, CI, artifact checks, and release checklist are available on GitHub.
-**Current status: preparation only.** No Qwen4 model weights, benchmark results, or validated Qwen4 architecture compatibility are available yet. Any future HF checkpoint release requires confirmation of the official source and license, GPU inference tests, and quality evaluation. This is an independent community project, not affiliated with Qwen or NVIDIA.
-- [HF preparation repository — no weights yet](https://huggingface.co/IteraCompute/Qwen4-27B-NVFP4)
-- [GitHub project](https://github.com/IteraCompute/qwen4-27b-nvfp4)
-- [Quantization and release checklist](https://github.com/IteraCompute/qwen4-27b-nvfp4/blob/main/docs/RELEASE.md)
-- [Release progress](https://github.com/IteraCompute/qwen4-27b-nvfp4/issues/2)
-中文：目前已开源量化工具和发布流程；模型权重将在官方模型发布并完成 GPU 与质量验证后准备上线。
-## Hugging Face Inference Provider
-Our Hugging Face Inference Provider integration is currently being prepared for review. Availability through HF will be announced after deployment and Hugging Face validation are complete. The catalogue above describes direct IteraCompute API access.
-## Resources
-- [Website](https://iteracompute.com/)
-- [Current models](https://iteracompute.com/models/)
-- [Free API and allowance](https://iteracompute.com/free/)
-- [GitHub organization](https://github.com/IteraCompute)
-- [API documentation](https://iteracompute.com/docs.html)
-- [Privacy policy](https://iteracompute.com/privacy.html)
-- Provider partnerships: [partnerships@iteracompute.com](mailto:partnerships@iteracompute.com)
-## Privacy
-Prompts and generated outputs are not retained by default. Operational metadata is processed only as needed for authentication, metering, reliability, abuse prevention, and billing. See our [privacy policy](https://iteracompute.com/privacy.html) for details.

+---
+title: README
+emoji: 🦀
+colorFrom: purple
+colorTo: pink
+sdk: static
+pinned: false
+---
+# IteraCompute
+**OpenAI-compatible inference for open-weight models. Start with free Qwen3.8-27B.**
+IteraCompute Artificial Intelligence Technology (Chongqing) Co., Ltd. provides reliable, usage-based inference for language models through OpenAI-compatible APIs. We focus on predictable integration, streaming responses, tool use, structured outputs, and privacy-conscious processing.
+## Start with free Qwen3.8-27B
+**Qwen3.8-27B is available for free through IteraCompute's OpenAI-compatible API.** Input, cached input, and output are all **$0 per 1M tokens**.
+- API model ID: `qwen/qwen3.8-27b`
+- API base URL: `https://api.iteracompute.com/v1`
+- Free tier: up to **30M tokens per account per UTC calendar month**, **128K total context**, and **160,000 TPM**.
+- Supports Chat Completions, streaming, and tool calling. Registration and email verification are required; no card or recharge is needed for the Free tier.
+The monthly allowance counts input and output, including reasoning, across the account's API keys. Free capacity is shared and may queue at busy times; fair-use limits apply and there is no uptime or latency SLA. See the current [free allowance and limits](https://iteracompute.com/free/) and [Free Tier Terms](https://iteracompute.com/free-tier-terms/).
+[Start free](https://iteracompute.com/free/) · [First API call](https://iteracompute.com/docs.html) · [Model details](https://iteracompute.com/models/qwen3-8-27b/)
+中文：Qwen3.8-27B 目前免费，输入、缓存输入和输出均为 $0；免费档每月最高 30M tokens、128K 上下文、160K TPM，无需信用卡或充值，高峰期可能排队。具体额度与可用性以官网和控制台为准。
+## Current model API catalogue
+The [website model catalogue](https://iteracompute.com/models/) lists the following models as online, checked on **2026-10-14**. These are direct IteraCompute API IDs; use them exactly as shown. Current availability, capabilities, and pricing are published on the website and in the Console.
+| Model | API model ID | Access |
+| --- | --- | --- |
+| [Qwen3.8-27B](https://iteracompute.com/models/qwen3-8-27b/) | `qwen/qwen3.8-27b` | Free within published allowance |
+| [Qwen3.8-2.4T-A95B](https://iteracompute.com/models/qwen3-8-2-4t-a95b/) | `qwen/qwen3.8-2.4t-a95b` | Usage-based |
+| [GLM-5.3-Flash](https://iteracompute.com/models/glm-5-3-flash/) | `z-ai/glm-5.3-flash` | Usage-based |
+| [Kimi K2.7 Code](https://iteracompute.com/models/kimi-k2-7-code/) | `moonshotai/kimi-k2.7-code` | Usage-based |
+| [DeepSeek V4 Pro](https://iteracompute.com/models/deepseek-v4-pro/) | `deepseek/deepseek-v4-pro-0813` | Usage-based |
+| [GLM-5.3](https://iteracompute.com/models/glm-5-3/) | `z-ai/glm-5.3` | Usage-based |
+| [Kimi K3](https://iteracompute.com/models/kimi-k3/) | `moonshotai/kimi-k3` | Usage-based |
+| [MiniMax M3](https://iteracompute.com/models/minimax-m3/) | `minimax/minimax-m3` | Usage-based |
+| [DeepSeek V4.1 Flash](https://iteracompute.com/models/deepseek-v4-1-flash/) | `deepseek/deepseek-v4.1-flash` | Usage-based |
+| [DeepSeek V4 Flash 0731](https://iteracompute.com/models/deepseek-v4-flash-0731/) | `deepseek/deepseek-v4-flash-0731` | Legacy alias for V4.1 Flash |
+[Current pricing](https://iteracompute.com/pricing/) · [Public Models endpoint](https://api.iteracompute.com/v1/models)
+## Open-source quantization
+### Qwen3.8-27B NVFP4 — public checkpoint
+[Download IteraCompute/Qwen3.8-27B-NVFP4](https://huggingface.co/IteraCompute/Qwen3.8-27B-NVFP4): the unchanged NVFP4 checkpoint deployed by IteraCompute on RTX 5090, with source provenance, Apache 2.0 license, SHA-256 checksums, and vLLM deployment instructions. Original model: Qwen. NVFP4 quantization: Gittensor Model Hub.
+We plan to upload a Qwen 4 NVFP4 version as soon as possible after official availability and licensing, quantization, and inference validation. Follow the [Qwen 4 preparation repository](https://huggingface.co/IteraCompute/Qwen4-27B-NVFP4); no Qwen 4 weights are available there yet.
+中文：Qwen3.8-27B NVFP4 检查点现已公开托管，保留上游来源、许可证及校验文件。我们后续也会第一时间推进并上传经过验证的 Qwen 4 NVFP4 版本，预告仓库已准备好。
+### Qwen4-27B NVFP4 — preparation stage
+We are preparing a reproducible NVFP4 quantization and release workflow for the anticipated Qwen4-27B model. The toolkit, CI, artifact checks, and release checklist are available on GitHub.
+**Current status: preparation only.** No Qwen4 model weights, benchmark results, or validated Qwen4 architecture compatibility are available yet. Any future HF checkpoint release requires confirmation of the official source and license, GPU inference tests, and quality evaluation. This is an independent community project, not affiliated with Qwen or NVIDIA.
+- [HF preparation repository — no weights yet](https://huggingface.co/IteraCompute/Qwen4-27B-NVFP4)
+- [GitHub project](https://github.com/IteraCompute/qwen4-27b-nvfp4)
+- [Quantization and release checklist](https://github.com/IteraCompute/qwen4-27b-nvfp4/blob/main/docs/RELEASE.md)
+- [Release progress](https://github.com/IteraCompute/qwen4-27b-nvfp4/issues/2)
+中文：目前已开源量化工具和发布流程；模型权重将在官方模型发布并完成 GPU 与质量验证后准备上线。
+## Hugging Face Inference Provider
+Our Hugging Face Inference Provider integration is currently being prepared for review. Availability through HF will be announced after deployment and Hugging Face validation are complete. The catalogue above describes direct IteraCompute API access.
+## Resources
+- [Website](https://iteracompute.com/)
+- [Current models](https://iteracompute.com/models/)
+- [Free API and allowance](https://iteracompute.com/free/)
+- [GitHub organization](https://github.com/IteraCompute)
+- [API documentation](https://iteracompute.com/docs.html)
+- [Privacy policy](https://iteracompute.com/privacy.html)
+- Provider partnerships: [partnerships@iteracompute.com](mailto:partnerships@iteracompute.com)
+## Privacy
+Prompts and generated outputs are not retained by default. Operational metadata is processed only as needed for authentication, metering, reliability, abuse prevention, and billing. See our [privacy policy](https://iteracompute.com/privacy.html) for details.