First look: Qwen 3.8-Max on Token Station

Alibaba previewed Qwen 3.8-Max at the World AI Conference on July 19, 2026: 2.4 trillion parameters, a sparse Mixture-of-Experts, native multimodality, and a claim that it trails only Claude Fable 5. What it does not have yet is a benchmark table, a model card, or a confirmed context window.

Rather than repeat the claim, we ran a small comparison. We sent the same three tasks through bailian-intl/qwen3.8-max-preview, kimi/kimi-k3, and openai/gpt-5.6 on Token Station, one key, temperature 0. Two are coding tasks with checkable answers; one is a probability problem.

The tasks

• Trapping rain water. A two-pointer problem; the answer is verified against known inputs.
• Probability. Three balls drawn without replacement from 3 red, 4 blue, and 5 green. Probability all three are different colors. Exact answer: 3/11.
• Find the bug. A merge_sorted function missing one line; the fix must pass the given test cases.

Each answer was checked automatically. The code was extracted and run; the fraction was matched.

Results

All nine passed. On tasks this small, that is a capability floor: every frontier model here solves them. The interesting signal is in the latency and token counts.

Qwen 3.8-Max was the most consistent. It finished every task in 7 to 8 seconds and used 142 to 165 reasoning tokens each time, regardless of the problem. GPT-5.6 was the fastest and tersest on the coding tasks (4 to 5 seconds, under 120 output tokens) but took 13.5 seconds on the probability question. Kimi K3 spent the least reasoning on the easy tasks (55 to 86 tokens) then ramped hard on the debugging task: 294 reasoning tokens and 31.5 seconds, where the other two stayed flat.

The honest read: on a small, automatically checkable set, Qwen 3.8-Max, still in preview, held its own with two shipping frontier models, and did it with the steadiest latency. That is a reason to point your own workload at it, not a verdict.

What is still unconfirmed

The test above is ours. The rest of the spec is Alibaba’s, and parts of it are still claims.

• 2.4 trillion parameters, sparse MoE. Active parameters per token are not disclosed, which is the number that determines serving cost.
• Multimodal. Text and visual inputs are confirmed. The full modality list (video, documents, speech, image generation) is not, with no spec sheet published.
• Context window: not published.
• “Second only to Fable 5.” Alibaba’s own words, from internal evaluations. No benchmark table or model card exists. For calibration, the predecessor Qwen 3.7-Max scored 92.4 on GPQA Diamond, 80.4 on SWE-bench Verified, and 69.7 on Terminal-Bench 2.0.
• Open weights. Alibaba says Qwen 3.8 is “going open-weight soon,” with no date or license. The Max tier has been closed so far; the open-weight line continued separately through Qwen 3.6.

At 2.4 trillion parameters it is the second-largest publicly known model, behind Moonshot’s Kimi K3 at 2.8 trillion, which shipped open-weight the same week.

Try it on Token Station

The endpoint is the standard Token Station OpenAI-compatible API:

curl https://models.bytefuture.ai/v1/chat/completions \
  -H "Authorization: Bearer TOKEN_STATION_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bailian-intl/qwen3.8-max-preview",
    "messages": [
      {"role": "user", "content": "Refactor this function and explain the change."}
    ]
  }'

Token Station also exposes the Anthropic API shape, so the same route drops into Claude Code.

Route it in your coding agent

In Claude Code, put it in the Opus slot:

# Token Station endpoint + auth
export ANTHROPIC_BASE_URL="https://models.bytefuture.ai"
export ANTHROPIC_AUTH_TOKEN="gw-YOUR_TOKEN_STATION_KEY"

export ANTHROPIC_DEFAULT_OPUS_MODEL="bailian-intl/qwen3.8-max-preview"
export ANTHROPIC_DEFAULT_SONNET_MODEL="bailian-intl/qwen3.8-max-preview"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="bailian-intl/qwen3.8-max-preview"
export CLAUDE_CODE_SUBAGENT_MODEL="bailian-intl/qwen3.8-max-preview"

claude

In Codex, make it the default model:

mkdir -p ~/.codex
cat > ~/.codex/config.toml <<'EOF'
model = "bailian-intl/qwen3.8-max-preview"
model_provider = "token_station"

[model_providers.token_station]
name = "token_station"
base_url = "https://models.bytefuture.ai/v1"
env_key = "TOKEN_STATION_API_KEY"
wire_api = "responses"
EOF

export TOKEN_STATION_API_KEY="gw-YOUR_TOKEN_STATION_KEY"
codex

Pricing

Qwen 3.8-Max is in preview and Alibaba has not published standard API pricing. Token Station passes the provider’s rate through at zero markup; check the dashboard for the current per-million rate. The $1 signup credit is enough to run a first evaluation.

The essentials:

• Base URL (OpenAI-compatible): https://models.bytefuture.ai/v1
• Base URL (Anthropic-compatible): https://models.bytefuture.ai
• Model: bailian-intl/qwen3.8-max-preview
• API key: starts with gw-, from the Token Station dashboard

Try Token Station