Qwen 3.8 Max, DeepSeek V4 Flash 0731 and MiniMax H3 (Hailuo 3) are live on 0G

At 02:15 UTC this morning, Alibaba posted the Qwen 3.8 Max announcement. At 04:56 UTC, the model was live on the 0G Private Computer. 2 hours 41 minutes from announce tweet to a working, attested endpoint, and you do not have to take our word for it: the 0G Router publishes a created timestamp for every model in its catalog, so the receipt is one curl away.

That was one of three arrivals in a single morning. Qwen 3.8 Max, Alibaba's 2.4-trillion-parameter flagship, joins on Verified Routing. DeepSeek V4 Flash 0731, the official release of the model half the timeline has been running since Friday, swapped into the existing deepseek-v4-flash id, so everyone already using it got upgraded without changing a thing. And MiniMax H3, the model behind Hailuo 3, becomes the first video generation model on the Private Computer, the same day its weights went public on Hugging Face. This post updates the running record of every model on 0G.

Three models, one morning

Qwen 3.8 Max: the 2.4T flagship, two hours later

Qwen 3.8 Max is Alibaba's new flagship, a 2.4-trillion-parameter Mixture of Experts model the company calls its most capable to date. It takes text, image, and video input, returns text, and carries a 1M-token context window (991K max input) with up to 128K tokens of output and a 256K-token reasoning chain. Alibaba's launch demo is a coding agent that ran 10+ days of self-directed development, from an empty folder to a production CLI tool, with the full project trace public on GitHub.

Per Alibaba's published results, the model tops its comparison chart on PaperBench (93.0) and IFBench (82.8) and posts 86.6 on Terminal Bench 2.1. The sharpest jump over Qwen 3.7 Max is agentic coding: DeepSWE 1.1 moves from 21.6 to 56.6 generation over generation. Alibaba says open weights follow next week. Within hours of the announcement, the timeline filled with head-to-head tests against Kimi K3 and the GPT-5.6 series; both of those are also reachable on 0G, so you can run the same comparison yourself on one endpoint.

On 0G it runs through Verified Routing, a TeeTLS-attested path straight to Alibaba's official endpoint. 0G attests the route; the inference runs at the upstream under its own data policy. It is the fifth Qwen model on the platform, joining Qwen 3.7 Max and the wider family that came out of the Alibaba Cloud partnership.

DeepSeek V4 Flash 0731: the upgrade nobody had to do

DeepSeek shipped V4 Flash to its official API on July 31, and the release has carried the timeline since: per DeepSeek's published chart, the 0731 build scores 82.7 on Terminal Bench 2.1 (its own preview scored 61.8), 54.4 on DeepSWE (7.4x the preview), and 25.2 on Agents' Last Exam. It is a light MoE, 284B total parameters with 13B active, with a native 1M-token context and up to 384K tokens of output. DeepSeek also made it speak the Responses API format natively and adapted it for Codex.

The reaction has been less about any single score and more about the math. Near-frontier agent benchmarks from a model with 13B active parameters changes how you budget: at $0.138 per million input tokens, rationing an agent's turns stops making sense. Within days the community had quantized builds small enough to run on high-end laptops.

On 0G, this launch is quieter by design. V4 Flash was already in the catalog, so the official release swapped in place: same deepseek-v4-flash id, same key, same Verified Routing tier, same price. If you called it yesterday, you are calling the 0731 build now. Zero action needed, and the catalog entry states the build directly: the model is served as the pinned 2026-07-31 snapshot. It sits next to DeepSeek V4 Pro on the same Verified Routing tier.

MiniMax H3: the first video model on the Private Computer

MiniMax announced H3 on July 31 as an omni-modal generation model and made the weights public on Hugging Face this morning at 02:39 UTC. Per MiniMax, it generates video at up to 2K resolution and 15 seconds per clip. Artificial Analysis's day-of-launch evaluation placed it #2 in text-to-video and #3 in image-to-video, the two modes 0G exposes at launch, and projected that an open-weights release would make it the strongest open-weights video model by far. This morning's weights drop delivered exactly that. By 13:23 UTC, day one of the weights being public, it was in the 0G catalog.

H3 is the first model on the Private Computer that outputs video; until today the catalog generated text, images, and transcripts. The launch scope on 0G is text-to-video and image-to-video from a first frame, at the 2K tier, billed at $0.195 per generated second. It runs on Verified Routing, with MiniMax's official API as the attested upstream, under the catalog id minimax-h3.

It is the second MiniMax model on the platform: MiniMax M3 has been serving multimodal text inference on Verified Routing since June.

Check the timestamps yourself

curl -s https://router-api.0g.ai/v1/models | jq '.data[] | select(.id=="qwen3.8-max").created'
# 1785732984  ->  August 3, 2026, 04:56:24 UTC

Alibaba's announce tweet is stamped 02:15 UTC; the gap is 2 hours 41 minutes, and both ends are public. The same query shows deepseek-v4-flash re-created at 07:41 UTC this morning, which is the in-place swap landing.

Where they sit on the trust tiers

A recurring line in the discourse around these releases: open weights and private are not the same thing. If you call a hosted API, your data leaves the building, and the standard community answer is to self-host. That answer works when you have the hardware; at 2.4 trillion parameters, Qwen 3.8 Max makes self-hosting a data-center project, so for most teams the hosted API is the only realistic path. The trust tiers exist for exactly that gap.

The Private Computer labels every model by what you can verify, not just what it does (trust tier docs). Private Inference (TeeML) seals the inference inside a 0G-operated enclave. Verified Routing (TeeTLS) attests the path to the provider's official endpoint, provable onchain, with inference running at the upstream. Standard Routing is the model through 0G's endpoint with no attestation.

Qwen 3.8 Max and DeepSeek V4 Flash 0731 both run on Verified Routing: the route from 0G's broker to the upstream is hardware-attested and provable onchain. For Qwen 3.8 Max that upstream is Alibaba's own cloud endpoint; for V4 Flash it is the providers serving the pinned 0731 snapshot; for MiniMax H3 it is MiniMax's official API. The full tier-by-tier catalog lives in the running record.

Build with it

Any OpenAI-compatible client reaches the new text models with one base URL change:

from openai import OpenAI

client = OpenAI(base_url="https://router-api.0g.ai/v1", api_key="app-sk-<YOUR_KEY>")
r = client.chat.completions.create(model="qwen3.8-max",  # or deepseek-v4-flash
    messages=[{"role": "user", "content": "Hello"}])
print(r.choices[0].message.content)

Existing V4 Flash integrations need no edit at all; the model id did not change. For Verified Routing models, set verify_tee: true in the request body; the response carries the attestation in x_0g_trace.tee_verified. MiniMax H3 runs on the same router through an async video flow: POST a job to /v1/videos with model minimax-h3, a prompt, and optionally a first-frame image, then poll the job and download the MP4 when it completes.

Pricing is USD-denominated and settled in $0G: Qwen 3.8 Max at $1.65 per million input tokens and $4.951 per million output, DeepSeek V4 Flash at $0.138 and $0.275, and MiniMax H3 at $0.195 per generated second of video. Exact per-model prices are on each model's page at pc.0g.ai.

Frequently asked questions

What just went live on the 0G Private Computer?

Three models on August 3, 2026: Qwen 3.8 Max (Alibaba's 2.4T flagship) and DeepSeek V4 Flash 0731, both on Verified Routing, and MiniMax H3 (Hailuo 3), the first video generation model on the platform.

Do existing DeepSeek V4 Flash users need to do anything?

No. The 0731 official release swapped into the existing deepseek-v4-flash id. Same key, same endpoint, same price; every existing integration was upgraded in place.

Is Qwen 3.8 Max on 0G private or verified?

Verified. It runs on Verified Routing (TeeTLS): the path to Alibaba's official endpoint is hardware-attested and provable onchain. The inference itself runs at Alibaba under its own data policy; this tier attests the route, it does not seal the inference.

What can MiniMax H3 do on 0G?

At launch, text-to-video and image-to-video from a first frame: clips of 4 to 15 seconds at the 2K tier, $0.195 per generated second, through the router's async /v1/videos endpoint.

How fast do new models arrive on 0G?

Hours, and you can check. Alibaba announced Qwen 3.8 Max at 02:15 UTC on August 3; the 0G Router's public created field stamps it live at 04:56 UTC, 2 hours 41 minutes later.

How do I call the new models?

Point any OpenAI-compatible client at https://router-api.0g.ai/v1 with an app-sk-* key from pc.0g.ai, and set the model id to qwen3.8-max or deepseek-v4-flash. For MiniMax H3, submit an async job to the same router's /v1/videos endpoint with model minimax-h3.

Build on 0G

• Try the new models now: pc.0g.ai
• Chat with models privately, no key: app.0g.ai/private-chat
• Read the trust-model details: docs.0g.ai
• Get $0G: get.0g.ai
• Follow @0G_labs for what goes live next on 0G private compute

Sources:

• 0G Router model list (live model ids, created timestamps, trust tiers, pricing)
• Alibaba's Qwen 3.8 Max announcement and Qwen blog (2.4T MoE, published benchmarks, open-weights timeline)
• DeepSeek's V4 Flash announcement and DeepSeek API docs (0731 release, published benchmarks, Responses API)
• MiniMax's H3 public availability post and MiniMax-H3 weights on Hugging Face (open weights, model capabilities)
• 0G router privacy and trust modes (Private / Verified / Standard definitions)
• The running record of every model on 0G (full catalog by trust tier)