Baltor Get started

Models

Qwen3.8-2.4T-A95B

By Qwen. 2.45 trillion parameters, a context window of 262,144 tokens and the licence other.

Facts

Released
2026-08-08 the day the repository was first published on Hugging Face Hugging Face, read
Licence
other the licence the model card declares Hugging Face, read
Open weights
Yes the weights are published in this Hugging Face repository Hugging Face, read
Parameters
2.45 trillion counted from the safetensors weight files Hugging Face, read
Active parameters
121.7 billion estimate estimated from the expert counts and sizes in the configuration model configuration, read
Knowledge cutoff
Unknown
Context window
  • 262,144 tokens the maximum position embeddings in the model configuration model configuration, read
Longest output
Unknown
Tool calling
  • Yes the chat template in the tokenizer configuration accepts tool definitions Hugging Face, read
Structured output
Unknown
Reasoning controls
Unknown
Inputs and outputs
text in, text out Hugging Face, read
Good for
No source names a use.

Prices

Each provider's own price in US dollars per million tokens, as an openly licensed source lists it
ProviderInputOutputContextAs ofSource
IteraCompute1.95 USD5.70 USD1,048,576 older than 30 days; check the providermodels.dev, read
Deep Infra2.00 USD6.00 USD262,144 older than 30 days; check the providermodels.dev, read
Eden AI2.00 USD6.00 USD1,000,000 older than 30 days; check the providermodels.dev, read
Kilo Gateway2.00 USD6.00 USD1,000,000 older than 30 days; check the providermodels.dev, read
Modal2.00 USD6.00 USD1,010,000 older than 30 days; check the providermodels.dev, read
Novita AI2.00 USD6.00 USD1,000,000 older than 30 days; check the providermodels.dev, read
OpenRouter2.00 USD6.00 USD1,048,576 older than 30 days; check the providermodels.dev, read
SiliconFlow2.00 USD6.00 USD1,049,000 older than 30 days; check the providermodels.dev, read
Together AI2.00 USD6.00 USD1,010,000LiteLLM, read
TensorX2.50 USD6.00 USD262,144 older than 30 days; check the providermodels.dev, read
Hugging Face2.50 USD6.25 USD262,144 older than 30 days; check the providermodels.dev, read
Merge Gateway2.50 USD6.25 USD262,144 older than 30 days; check the providermodels.dev, read

Run it on your own hardware

Weights are the sizes of the files a source lists, or an estimate from the parameter count where none does. Memory is an estimate: weights plus KV cache plus 512 MiB and 5 percent of the weights for runtime buffers. Check it against your hardware.

QuantizationWeightsMemory at 8,192 tokensMemory at 32,768 tokens
UD-IQ1_S473 GiB498 GiB501 GiB
UD-IQ1_M525 GiB553 GiB555 GiB
UD-IQ2_XXS611 GiB643 GiB645 GiB
UD-IQ2_XS680 GiB716 GiB718 GiB
UD-IQ3_XXS890 GiB936 GiB938 GiB
UD-IQ4_XS1,221 GiB1,283 GiB1,285 GiB
Q8_02,422 GiB2,544 GiB2,546 GiB
BF164,557 GiB4,786 GiB4,788 GiB

Start it with a local runtime

  • Ollama
    ollama run hf.co/unsloth/Qwen3.8-2.4T-A95B-GGUF:UD-IQ1_S
  • LM Studio
    lms get unsloth/Qwen3.8-2.4T-A95B-GGUF
    lms server start
  • llama.cpp server
    llama-server -hf unsloth/Qwen3.8-2.4T-A95B-GGUF:UD-IQ1_S --jinja
  • vLLM
    vllm serve Qwen/Qwen3.8-2.4T-A95B
  • SGLang
    sglang serve --model-path Qwen/Qwen3.8-2.4T-A95B --port 30000

Use it from your harness

Set up for DeepInfra with the model Qwen/Qwen3.8-2.4T-A95B. Each endpoint page has the same setup for its own address.

OpenCode

Put this in opencode.json in your project folder:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "deepinfra": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "DeepInfra",
      "options": {
        "baseURL": "https://api.deepinfra.com/v1/openai",
        "apiKey": "{env:DEEPINFRA_TOKEN}"
      },
      "models": {
        "Qwen/Qwen3.8-2.4T-A95B": {
          "name": "Qwen3.8-2.4T-A95B"
        }
      }
    }
  }
}
  • OpenCode reads any OpenAI-compatible address through the @ai-sdk/openai-compatible package, and an address that speaks the Responses API through @ai-sdk/openai.

From OpenCode documentation, read .

Pi

Put this in ~/.pi/agent/models.json:

{
  "providers": {
    "deepinfra": {
      "baseUrl": "https://api.deepinfra.com/v1/openai",
      "api": "openai-completions",
      "apiKey": "$DEEPINFRA_TOKEN",
      "models": [
        {
          "id": "Qwen/Qwen3.8-2.4T-A95B"
        }
      ]
    }
  }
}
  • The apiKey field can name an environment variable as $NAME.

From Pi documentation, read .

Codex

Codex speaks only the Responses API, and DeepInfra documents no Responses address. A gateway that offers one can sit in between.

  • Codex speaks the Responses API only: responses is the one supported wire API of a custom provider. Ollama and LM Studio are built in and start with --oss.

From Codex documentation, read .

Claude Code

Put this in ~/.claude/settings.json, or variables in your shell:

export ANTHROPIC_BASE_URL="https://api.deepinfra.com/anthropic"
export ANTHROPIC_AUTH_TOKEN="$DEEPINFRA_TOKEN"
export ANTHROPIC_MODEL="Qwen/Qwen3.8-2.4T-A95B"
claude
  • Claude Code sends Anthropic Messages requests to ANTHROPIC_BASE_URL. Anthropic says it does not support routing Claude Code to models other than Claude through any gateway, so some features may not work with another model.

From Claude Code documentation, read .

Other ways to reach it

Published results

These are results other people published. Baltor did not run them and does not rank models by them.

Two more publishers are linked, never copied, because their terms do not allow it:

  • Artificial Analysis publishes its own intelligence, speed and price measurements.
  • OpenRouter lists current prices and uptime for each provider it routes to.