OpenChat

C-RLFT-tuned open 7B that once matched ChatGPT on benchmarks

7B C-Rlft Commercial Use Conversational Efficient Open Source

LLMs & Chat Open Source Has API Open Source

Researched 8 May 2026, 20:44 SGT · Published 8 May 2026, 08:00 SGT · Reviewed 11 Jul 2026

Visit OpenChat Compare alternatives

RECATOOLS Score

6.3 / 10

Capability

Value for money

Ease of use

ASEAN readiness

API quality

Founded

2023

Beijing, China

Users

200k+ downloads

Launched

Jul 2023

Developer

Tsinghua University

Overview

Apache-2.0 open models from Tsinghua's Guan Wang, trained with C-RLFT to learn from mixed-quality data. The 7B OpenChat 3.5 matched early-2023 ChatGPT on several benchmarks, but the line has been quiet since OpenChat 3.6 in May 2024.

Pricing

Pricing shown for reference only. These figures reflect RECATOOLS research as of 11 Jul 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.

Free

Fully free

Use cases

Running a capable open-source chatbot on a single consumer GPU for development Building a production chatbot at lower inference cost than GPT-3.5 API Research into efficient model training using mixed quality training data

ASEAN Perspective

OpenChat in Southeast Asia

ASEAN-region availability and pricing notes coming soon. Drop the editorial team a note via /contact/ if you can supply local context (Singapore/Malaysia/Indonesia/Thailand/Vietnam).

RECATOOLS Verdict

OpenChat earned its reputation honestly: the 7B OpenChat 3.5 traded blows with early-2023 ChatGPT on public benchmarks thanks to C-RLFT, a training method that squeezes signal out of mixed-quality data, all under Apache 2.0. The problem is the calendar. The last release was OpenChat 3.6 on Llama 3 8B in May 2024, and the project has been effectively dormant since — GitHub issues sit unanswered and no new model has followed. Qwen and Llama's newer small models have passed it on reasoning, multilingual coverage and context length. It still runs fine as a lightweight local model, but in 2026 there's no reason to start a new project on it. A significant open-LLM chapter, not a current recommendation.

Independent AI-assisted assessment by RECATOOLS.

What people say

The last OpenChat release was May 2024. That's the single most important fact about it in 2026: OpenChat 3.6, a Llama 3 8B fine-tune, arrived, the GitHub issues started piling up unanswered, and the project has been effectively quiet since.

Which is a shame, because the underlying work was genuinely clever. Guan Wang and collaborators at Tsinghua built C-RLFT, a training method that learns from mixed-quality data — treating expert responses and mediocre ones as differently conditioned examples instead of discarding the mediocre — without needing preference labels. The result, OpenChat 3.5, was a 7B Mistral fine-tune that traded blows with the March 2023 version of ChatGPT on public benchmarks; the January 2024 '0106' update went further, with the README cheerfully noting it beat Grok-1, a 314B model, on three of four benchmarks. All of it Apache 2.0, all of it runnable on a consumer GPU.

The field didn't stand still. Qwen 2.5 and its successors, Llama 3.x and Mistral's own instruct models now beat OpenChat at the same sizes, with longer context, better multilingual coverage and active maintenance. There's no first-party API and never was; you self-host or lean on a community endpoint.

If you already run OpenChat somewhere, it won't fall over. But nobody should start a new deployment on a model whose upstream went dark two years ago. File it with Zephyr and Vicuna: an important chapter in how open models learned to punch up, superseded by the models it helped make possible.

Summary of public user & expert reviews, compiled by RECATOOLS.

Notable facts

OpenChat 3.5 matched GPT-3.5 performance on multiple benchmarks with only 7 billion parameters — an 8x parameter efficiency advantage.
The C-RLFT training technique was invented in a university lab as a research project, then open-sourced — demonstrating that frontier training techniques can emerge from academic settings.
OpenChat was one of the first models to demonstrate that learning from negative examples (suboptimal responses) is as important as learning from positive examples.

Frequently asked questions

Is OpenChat free?

Yes. Apache 2.0 licence — free and commercial use permitted.

How does OpenChat compare to Llama?

OpenChat is fine-tuned on conversational data and achieves better instruction following than base Llama models.

What is C-RLFT?

Conditioned Reinforcement Learning Fine-Tuning — a technique that trains models to distinguish high-quality from suboptimal responses.

Can OpenChat run on a consumer GPU?

Yes. The 7B model runs on 8GB VRAM.

Is OpenChat better than Mistral 7B?

On conversational benchmarks, OpenChat often outperforms Mistral 7B due to better instruction tuning.

About this listing

Researched on Friday, 8 May 2026 at 20:44 SGT (UTC+8)

Published on Friday, 8 May 2026 at 08:00 SGT (UTC+8)

Last reviewed Saturday, 11 July 2026 (1 week ago)

This entry was compiled from publicly available data including OpenChat's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with OpenChat unless explicitly stated.

Data accuracy

Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.

For the latest details, please refer to OpenChat directly →

Spotted something out of date? Suggest an update →