Qwen 3.8 Max 2.4T Open Weights: 86.6% Agents [2026]

Qwen3.8 Max is Alibaba 2.4T total with 95B active per token across 512 experts with 10 routed plus 1 shared. Open weights Qwen3.8-2.4T-A95B went live August 12 2026 on ModelScope, while hosted qwen3.8-max API offers multimodal 1M context at $2 input and $6 output with $0.25 cache.

• First Max open weights: prior Max tiers stayed API-only, A95B is downloadable base.
• Agent benchmark jumps: 86.6 Terminal 2.1 versus 74.5 for 3.7 Max, 87.3 SWE-bench.
• API vs weights differ: API is multimodal with tools, weights are text-only thinking-on.

Why 95B active of 2.4T matters

Sparse MoE fires 4% per token, so 2.4T capacity costs 95B compute. Hybrid backbone mixes Gated DeltaNet linear attention with Gated Attention every fourth layer over 92 layers, trained with Multi-Token Prediction. Native context is 256K, expandable to 1.01M with undocumented quality trade-offs.

In company test Max spent 16 days building oh-my-cli with 265 commits, 127 PRs, and 151 issues, plus 33 GPU rounds beating paper method by 2.7 on AIME24. Treat as vendor direction and replicate slice, as with Qwen fast inference lane where catalog speed needed p95 engineering.

Task -> Router
  |- Qwen Max API ($2, multimodal, tools) -> hard coding + cowork
  |- Qwen 27B (self-host, 256K) -> sovereign + edge
  |- DeepSeek Flash ($0.14) -> bulk default
       |
       v
Eval: Terminal + SWE + cost per merged PR

Route hardest long-horizon work to Max, interactive to 27B Cerebras, bulk to Flash per DeepSeek Codex playbook.

Benchmark table: coding, agents, cost

From Alibaba Aug 2026, Vals, Arena, Benchgen, DataCamp Sep 2.

Vals notes Max matches Opus 4.7 at 2.3x lower cost $2.68 versus $6.17 per test. September 0902 update lifts DeepSWE 56.6 to 69.3 and Terminal 3.0 11.3 to 29.0. *SWE-Pro 67.7 on vendor card versus 87.3 resolve on third-party slice; track harness.

Step 1: Call Max API with effort control

API supports OpenAI chat, Responses, and Anthropic interfaces with xhigh, medium, low effort and preserve_thinking default.

# file: setup.sh
python3.12 -m venv .venv && source .venv/bin/activate
pip install openai==1.99 anthropic==0.66 tiktoken==0.9
 export QWEN_API_KEY=qwen-xxx

# file: max_client.py
import os
from openai import OpenAI
client = OpenAI(base_url="https://dashscope.aliyuncs.com/compatible-mode/v1", api_key=os.environ["QWEN_API_KEY"])
resp = client.chat.completions.create(
  model="qwen3.8-max",
  max_tokens=3000,
  extra_body={"reasoning_effort": "xhigh", "preserve_thinking": True},
  messages=[{"role": "system", "content": "You are a senior coding agent. Return diffs only."},
            {"role": "user", "content": "Reproduce research pipeline stage with tests"}]
)
print(resp.choices[0].message.content[:3000])

Limits 2M TPM and 15K RPM flat across 1M window. Cache $0.25 implicit, $0.17 explicit read. Pin qwen3.8-max exact snapshot plus 0902 note.

Step 2: Understand weights versus API gap

Weights A95B are text-only with thinking required-on under custom Qwen3.8-Max License, not Apache. Commercial over 100M MAU or $20M monthly revenue must display model name and contact model-business. API adds image, video, document over 200 pages, 100-hour video, Qwen-MM-Plugins for Blender and CAD, and built-in tools.

# file: weights.sh
# ModelScope Aug 12, Hugging Face mirror
pip install modelscope
modelscope download --model Qwen/Qwen3.8-2.4T-A95B --local_dir ./qwen-max-aw95b
ls -lh ./qwen-max-aw95b | head
# serving needs vLLM or SGLang cluster, not single GPU

Self-host reality is 427x RTX 4090 24GB, 94x A100 80GB, or 98x M3 Max 128GB. For nearly every team API is smarter; use weights for research and distillation, not prod serving. Details align with Fable benchmark production guide on price per task over raw params.

Step 3: Build long-horizon coding loop

Max targets 10-plus-day autonomy with cowork memory and visual self-inspection. Checkpoint every super-step and verify with evals.

# file: loop.py
STEPS = ["intake", "implement", "test", "repair", "visual_check"]
BUDGET = {"tokens": 150000, "usd": 2.50, "steps": 40}

def run(task: str):
  usage = {"tokens": 0, "usd": 0.0, "steps": 0}
  for s in STEPS:
    # call Max API with effort xhigh for implement, medium for others
    usage["steps"] += 1
    if usage["usd"] > BUDGET["usd"]:
      return {"halt": True, "reason": "budget"}
  return {"done": True}

# file: bench.sh
python evals/run_terminal.py --model qwen3.8-max --set mini-20 --effort xhigh
python evals/run_swe.py --model qwen3.8-max --set hard-10

Require diff plus test log under 900 words, visual screenshot diff for frontend tasks referencing Arena consumer-product strength where Max ranks #2.

Production reality check and failure modes

Four traps erase Max gains. First, 256K native versus 1M expandable confusion overflows quality: default to 256K working set with file memory, expand only for docs over 200 pages. Second, thinking always-on bills unexpectedly: set xhigh only for implement, medium for triage. Third, weights without tools underperform API: add BYO tools and vision harness or stay on API. Fourth, license breach on commercial scale: track MAU and revenue triggers and display attribution before launch.

Add guardrails from Opus automation workflow: 40-step cap, $2.50 thread budget, Postgres checkpoints, human approve for prod writes, and OTel cost per merged PR. Measure $2.68 Vals cost as baseline and alert on 25 percent rise.

When to pick Max versus 27B versus Flash

Pick Max API for hardest coding and cowork at $2 with multimodal tools. Pick 27B for sovereign self-host and interactive 1500 tok/s fast lane. Pick DeepSeek Flash for bulk default at $0.14. Most enterprises run Max hard lane with 27B fallback for data-bound jobs.

Step 4: Cost ledger and migration from 3.7 Max

Ship weekly ledger joining model, effort, tokens, and merge rate. Block promo when cost per merged PR rises even if pass rate flat.

# file: ledger.py
import json
from collections import defaultdict
rows=[json.loads(l) for l in open("runs.jsonl")]
by=defaultdict(list)
for r in rows:
  by[r["model"]].append(r)
for m,rs in by.items():
  mg=[x for x in rs if x.get("merged")]
  cpp=sum(x["cost_usd"] for x in rs)/max(1,len(mg))
  print(f"{m}: {len(mg)}/{len(rs)} cpp ${cpp:.2f}")

Migrate from 3.7 Max in one sprint. First, swap model ID in staging and run Terminal mini twenty plus SWE hard ten side by side for three days. Second, set xhigh for implement and medium for triage to avoid always-on bill shock. Third, freeze prompts for cache and enable Postgres checkpoints. Fourth, cut over hard lane to Max with 27B fallback for sovereign jobs. Version model, effort, and prompts together for reproducible audits.

Keep Flash bulk lane at low effort so Max stays focused where 86.6 Terminal and 93 PaperBench pay. Review 0902 deltas monthly because coding lifts arrive via post-training without base change.

Pin 0902 snapshot, retain visual diffs, and publish price per task weekly for leadership.

Document native 256K versus expandable 1M trade-offs in runbook to prevent quality regressions.

Keep rollback commit tagged for instant revert during post-training behavior shifts.

By Dr. Aris Thorne, Lead AI Research Fellow at Daily AI World.

Last tested & verified: September 2026 with Python 3.12, QwenCloud API, ModelScope Aug 12 weights and Vals Sep 2026.