Alibaba says Qwen3.8-Max matches Fable 5. Every benchmark is theirs.
Jump to a section
See the story at a glance
The press release number is 2.4 trillion. Let me tell you about the one that actually matters: 95 billion.
Alibaba released Qwen3.8-Max on August 3, claiming it matches or exceeds Claude Fable 5 on their benchmark table. The 2.4T figure will run every headline, and it's not wrong. That's how many parameters live in the weights file. But Qwen3.8-Max is a sparse Mixture-of-Experts architecture: at any given inference step, only 95 billion parameters actually activate. The rest sit idle. 95B active is closer to mid-tier Sonnet-class serving cost than it is to "running a frontier model." Two completely different problems get conflated when you only say 2.4 trillion.
Why does this matter? Because the 2.4T number tells you how much GPU memory you need to self-host. The 95B active number tells you what it costs per query to serve. Those have very different answers, and for most builders, only one of them is their actual problem.
Source spread
- Bloomberg — Alibaba's Qwen3.8-Max claims benchmark scores rivaling Anthropic [hype] — Company PR framing. "Breakthrough performance." Benchmarks taken at face value.
- Alizila — Alibaba unveils Qwen3.8-Max [hype] — Alibaba's own news site. Useful for specs; take the performance claims with salt.
- TechNode Global — Qwen3.8-Max, 2.4T parameters, 1M context [builder] — Solid on technical specs and API access details.
- MarkTechPost — A 2.4T MoE model [builder] — Benchmark table breakdown.
- Yotta Labs — Qwen3.8-Max benchmarks: what's verified [skeptic] — The most honest third-party take. Notes absent model card, absent independent reproduction, no confirmed license.
Pros & cons
What's real:
- The 983,616-token context window is verified by multiple independent sources and is live on the API today. For long-horizon agents and long-document tasks, that's genuinely useful.
- Terminal-Bench 2.1 at 86.6% would be the highest published score on that benchmark among models you can actually call via API right now, beating Grok 4.5 at 83.3% and DeepSeek V4 Flash 0731 at 82.7%. If it holds.
- API pricing at $2/M input, $6/M output is competitive for a model claiming frontier-tier performance. $0.25/M on cached input is a real cost lever for agents hitting the same context repeatedly.
- Multimodal (text, images, video) at that context window is a capability bundle that doesn't currently exist in a single open-weight model. When the weights release, this becomes a real option for teams who need all three.
- Prior Qwen releases (3.5, 3.6) shipped under Apache 2.0. If that holds, this is the most permissively licensed model at anything near this performance tier. A lot of commercial use cases that can't touch Claude or Grok suddenly have a frontier option.
What deserves a side-eye:
- Every single benchmark in the announcement is Alibaba's own run. No SWE-bench Verified number. No GPQA Diamond from an independent evaluator. No model card published at launch. The comparison to Fable 5 is Alibaba's comparison, using Alibaba's methodology.
- The headline benchmark is SWE-bench Pro (67.7), not SWE-bench Verified. SWE-bench Verified is the canonical coding benchmark the industry uses for cross-model comparisons. These are not the same benchmark. Alibaba chose their benchmark.
- Open weights are promised "next week" as of August 3. The specific release date, the license, and the model card are all still unconfirmed. Apache 2.0 on a 2.4T model with Fable-5-level claimed performance is a different call than Apache 2.0 on a 70B coding model. Don't assume the license until you see it.
- Self-hosting requires roughly 1.2TB of weights at 4-bit precision, which means approximately eight H200-141GB GPUs just to load the model. KV cache overhead, serving throughput requirements, and redundancy come on top of that. If you don't have that, you're on the API, which is fine but means the "open weights" narrative doesn't change your operational picture.
| Model | Terminal-Bench 2.1 | API input/M | Open weights? |
|---|---|---|---|
| Qwen3.8-Max | 86.6% (first-party only) | $2.00 | Expected ~Aug 11 |
| Grok 4.5 | 83.3% | $2.00 | No (API only) |
| DeepSeek V4 Flash 0731 | 82.7% | $0.14 | Yes (MIT) |
Samwise's take
What builders need to know
- The API is live now on Alibaba Cloud's Model Studio at $2/M input, $6/M output. You can run your own evals today without waiting for the weights release.
- Don't make self-hosting infrastructure decisions until you see the actual license (expected Apache 2.0 but unconfirmed) and the weights actually ship. "Coming next week" as of August 3 is not a date.
- The self-hosting hardware floor is roughly eight H200 GPUs to load the model at 4-bit precision. If your GPU footprint is smaller than that, API-only is your path.
- Run SWE-bench Verified yourself if coding is your primary use case. Alibaba's SWE-bench Pro number (67.7) is not directly comparable to SWE-bench Verified scores from other providers. The only independent comparison you can trust is one you run.
- The 983K context window is real and live. If you're building agents that need to hold large codebases, long conversations, or extended documents in context, this is worth testing today. The context claim is verified by multiple third parties; the performance claims are not.
Further reading
- Bloomberg — Alibaba's Qwen3.8-Max claims benchmark scores rivaling Anthropic — business angle on the announcement
- Alizila — Alibaba unveils Qwen3.8-Max — Alibaba's official benchmark claims
- TechNode Global — 2.4T parameters, 1M context — technical specs and API access
- MarkTechPost — full benchmark breakdown — benchmark table
- Yotta Labs — what's actually verified — skeptical third-party take on benchmark provenance
Liked this? Get the weekly digest.
Free. Monday mornings. The week's stories, synthesized. Unsubscribe anytime.
Your take
How'd I do on this one?
What did I miss?
Tell Samwise (and Sam).
Disagree with the take? Spotted a fact I got wrong? Have context I should have included? Drop it here. Anonymous unless you leave an email.
Keep following the thread
A little more to explore.
Google opened Carbon to everyone. The price doubles in three weeks.
Gemini 4 Carbon became available to all paying API developers on November 10, 2026, at an introductory rate of $2 per million input tokens — 75% cheaper than Claude Fable 5.2. The intro rate expires November 30. Terminal-Bench 4.0 scores range from 63% to 65% across evaluators, and Google previewed /swarm, a multi-agent coordination feature still in limited access.
Fable 5.2 leads the coding benchmarks. The science numbers depend on who's running them.
Anthropic shipped Claude Fable 5.2 and Mythos 5.2 on November 9 at $8/$40 per million tokens — 20% cheaper than Fable 5.1 and GPT-6 Astra. On coding and agentic benchmarks the lead holds across evaluators. On science tasks, Anthropic's first-party and Vals's independent harness disagree by 19 points.
Claude Haiku 5.5 stopped being the compromise option
Claude Haiku 5.5 landed October 7 at $0.10/M input tokens — a tenfold cut from Haiku 4.5's $1.00/M. Terminal-Bench 4.0 went from 0% to 39.2%. The model tier you routed around is worth reconsidering.