Atria Dawn vs Qwen 3.8 Max: The New Agentic Bet That Wants to Skip the 95B-Active Front Door

There is a specific number that makes this matchup worth writing down, and it is not any benchmark score. Qwen 3.8 Max, Alibaba​'s flagship released August 3, is a 2.4-trillion-parameter sparse MoE with roughly 95 billion active parameters per token, a 1M context, text/image/video input, and a $2.00/$6.00 rate card — the model that carried August's agentic conversation on independent leaderboards. Atria Dawn Preview, announced "now live" September 14 by the Shanghai AI Laboratory, is an 8-active-expert 744B-total MoE built on the GLM-5.2 foundation, MIT-licensed, 256K context, text-only, downloadable in BF16 and FP8 — a fraction of Qwen 3.8 Max's compute per token, aimed at the same long-horizon agentic buyer from the opposite end of the economics. The number worth writing down is the active-parameter gap: roughly 95B active versus 8 experts on a 744B base — and the question is whether a loop-first open model can beat a 95B-active frontier generalist by outworking it rather than outthinking it.

The sourcing asymmetry is severe and must stay visible. Qwen 3.8 Max's numbers are independent and checkable — an AA Intelligence Index of 40 (and an Agentic Index that has run as high as 58 on the old scoring), $2/$6, 1M context. Atria Dawn Preview's benchmark rows are vendor-reported from its model card, unreproduced by any neutral lab, and its international API at api.atria-asi.ai has no published price. On the card's grid — which compares Atria Dawn Preview against a field that includes Qwen 3.8 Max — Atria reports leading on discovery and tool-use suites (AutomationBench 53.8, BrowseComp 92.5, DeepSearchQA 96.0, BFCL v4 77.0, CyberGym 86.5) and trailing Qwen 3.8 Max's lineage on coding rows (SWE-bench Pro 59.6, Terminal-Bench 2.1 78.3). None of the Atria numbers is verified.

Two different bets on what "agentic" means

Qwen 3.8 Max is the bet that scale still wins. A 2.4T-total sparse MoE with ~95B active, a 1M window, native vision and video, native tool use, and an independent agentic footprint that made it the highest-scoring Chinese model of August — that is a front-door generalist: one very capable model that can be pointed at any agentic task and simply be smart enough to grind through it. It is priced at $2/$6, which for a frontier-class model is the aggressive part of its story, and it is served at that rate through the vendor's own API and several third-party platforms.

Atria Dawn Preview is the opposite bet: that the loop is the product and raw scale is a cost you should not pay. It is a 744B-total MoE with 8 experts active per token — a fraction of Qwen 3.8 Max's active compute — with a 256K text-only window, MIT weights, and a tuning whose four pillars (Discovery, Creation, Delivery, Cybersecurity) all describe the same behaviour: analyse, design, use tools, write and run code, read the experimental result, recover from failure, iterate. Its reported strengths are long-horizon suites; its reported weaknesses are the coding rows where a 95B-active generalist wins. The active-compute gap is the whole economic story: if the loop works, you get agentic execution at a fraction of Qwen 3.8 Max's hardware cost, on hardware you own.

• Active compute — Atria Dawn Preview 8 experts on 744B total vs Qwen 3.8 Max ~95B active on 2.4T total

• Price — Atria unpublished (international API), free weights vs Qwen 3.8 Max $2.00/$6.00 per 1M

• Context — Atria 256K text-only vs Qwen 3.8 Max 1M, text/image/video input

• Independent record — none for Atria vs Qwen 3.8 Max AA Index 40

Reading the grid without fooling yourself

The vendor-reported grid claims Atria Dawn Preview leads Qwen 3.8 Max on the agentic suites — AutomationBench 53.8, BrowseComp 92.5, DeepSearchQA 96.0, CyberGym 86.5 — while losing the coding rows. The split is coherent with the model's design: a loop-first agent should win tasks where it gets to iterate and lose tasks that reward raw single-shot reasoning. But every row is the lab's own harness, unreproduced, and the comparison includes no neutral ground at all. Qwen 3.8 Max's independent footprint is real and checkable today; Atria Dawn Preview's is entirely a self-report. The first neutral run will matter more than this entire grid.

For a team already running Qwen 3.8 Max, nothing here argues for switching — its 95B-active compute, 1M window, and verified agentic record are exactly why it holds the front door. For a team whose pain is the loop — the model stopping to ask, losing the thread on a long experiment — Atria Dawn Preview's open weights and loop-first tuning are aimed precisely there, and its eight-active-expert economics are the reason a self-host is even plausible. Both models sit behind one key on a routing platform like OrcaRouter — Qwen 3.8 Max at pass-through $2/$6 with zero markup and automatic failover, and Atria Dawn Preview routable the moment the Shanghai lab publishes a rate — so the only honest way to settle this matchup is to run both on your own workload and let the loop decide.

The verdict

Do not read the vendor grid as proof, and do not read the active-parameter gap as a promise — one column is independently verified, the other is a lab's pitch, and an 8-active-expert model beating a 95B-active flagship on long-horizon work is exactly the kind of claim that needs a neutral lab before anyone should trust it. If you want a proven agentic front door today, Qwen 3.8 Max is the model and $2/$6 is a fair price for it. If your bottleneck is the loop itself, Atria Dawn Preview is the most interesting experiment in this batch — open, cheap to run, and worth a controlled eval behind failover, with the understanding that its numbers will change the first time someone who does not work for the lab runs it.