Coding Horizon

Qwen 3.8 Max Is The Open Model Moment We Waited For

Every figure, name and benchmark this video renders on screen, chased to a source. Organised by the chapter that puts it on screen.

The model itself

Fact	Value	Source
Total parameters	2.4 trillion	Qwen3.8-2.4T-A95B model card
Active parameters per token	95 billion	Qwen3.8-2.4T-A95B model card
Routed experts	512	vLLM recipe, Qwen3.8-2.4T-A95B
Routed experts selected per token	10	Qwen3.8-2.4T-A95B model card
Shared experts	1, always on	Qwen3.8-2.4T-A95B model card
Layers	92 (23 full attention, 69 linear attention)	vLLM recipe
Native context	262,144 tokens	Qwen3.8-2.4T-A95B model card
Maximum context	1,010,000 tokens (1M class)	Qwen3.8-2.4T-A95B model card
Maximum output tokens	131,072	Qwen3.8-2.4T-A95B model card
Maximum thinking budget	262,144 tokens	Qwen3.8-2.4T-A95B model card
Announced	3 August 2026	MarkTechPost launch report
Ratio to Qwen3.8-27B	2.4T / 27B is 88.9x	arithmetic on the two model cards

The hosted product reads text, images and video and can call tools, search and run code. The downloadable checkpoint is text only and always reasons: thinking mode cannot be disabled and every response opens with a <think> block. Model card, DataCamp on the hosted model’s modalities.

The benchmark table

Qwen’s own launch comparison, reproduced by emergent.sh. These are vendor numbers run in Qwen’s harness, not independent reproductions, which is the caveat the video states out loud.

Benchmark	Qwen3.8-Max	Qwen3.7-Max	GPT-5.6 Sol	Claude Opus 4.8	Fable 5
Terminal Bench 2.1	86.6	74.5	88.8	84.6	84.6
SWE-bench Pro	67.7	60.6	64.6	69.2	80.0
PaperBench	93.0	64.8	90.5	80.3	88.8
GPQA Diamond	92.6	92.4	94.1	92.0	92.6

• Terminal Bench gain, 74.5 to 86.6, is 12.1 points.
• PaperBench gain, 64.8 to 93.0, is 28.2 points.
• JobBench rises 31.3 to 53.4. apidog benchmark breakdown
• DeepSWE is Qwen’s clear loss: 56.6, against Opus 4.8 at 59, Gemini 5 at 70 and GPT-5.6 Sol at 73. MindStudio, where it really ranks

Terminal Bench 2.1 places an agent inside a terminal and scores whether it finishes real engineering tasks under a time limit. PaperBench, built by OpenAI, scores whether an agent can reproduce a scientific paper’s results from its experimental description. emergent.sh

Pricing, international Qwen API

Model	Input, per 1M tokens	Output, per 1M tokens
Qwen3.7-Max	$2.50	$7.50
Qwen3.8-Max	$2.00	$6.00

Both input and output fall by exactly 20 percent. emergent.sh pricing table, DataCamp

The sixteen day agent

Alibaba’s headline demonstration: Qwen3.8-Max worked autonomously on the command line tool oh-my-cli for sixteen days, taking user requests, turning them into GitHub issues, assigning them to itself, writing the code, running the tests and iterating.

• By 30 July 2026: 265 commits, 127 pull requests, 151 issues, with no human commit.
• Two further runs: a research paper reproduction using roughly 125 hours of compute, 7,600 lines of code and 33 GPU training jobs; and a simulated e-commerce business that took 100,000 yuan to 416,252 yuan.

the-decoder, developer-tech

The ladder of open models

Model	Total params	Active params	Context	Terminal Bench 2.1
Qwen3.8-27B	27B (27.78B)	dense	262K, to 1M	73.0
Qwen3.8-Flash-Next	125B	6B	262K, to 1M	—
DeepSeek V4 Flash	284B	13B	1M	82.7
GLM-5.3-Flash	320B (321B)	18B	1M	84.3
Kimi K3	2.8T	104B	1M	88.3
Qwen3.8-Max	2.4T	95B	262K, to 1M	86.6

• Qwen3.8-27B: 73.0 Terminal Bench 2.1, 61.7 SWE-bench Pro. Qwen3.8-27B model card
• A good 4 bit build of the 27B lands around 17 GB of weights, which is the 24 to 32 GB class once context is allowed for. codersera, the 17GB open model
• Qwen3.8-Flash-Next: 125B total, 6B active, 262,144 native context extensible to 1M. DataCamp
• DeepSeek V4 Flash: 284B total, 13B active, 1M context. vLLM recipe
• GLM-5.3-Flash: 321B total, 18B active, 1M context. vLLM recipe
• Kimi K3: 2.8T total, 104B active, 1M context, 88.3 on Terminal Bench 2.1 in Moonshot’s own table. Moonshot K3 specs

Kimi’s 88.3 and Qwen’s 86.6 come from two different vendors’ harnesses, so the pair does not settle an ordering between them.

The hardware, from the published serving recipe

vLLM’s verified recipe for Qwen3.8-2.4T-A95B. recipes.vllm.ai, vLLM day 0 blog, 12 August 2026

Precision	Weights	B300 (268 GB)	H200 (141 GB)
BF16, 16 bit	4.45 TiB (4.89 TB)	24	48
FP8, 8 bit	2.27 TiB	16	32
NVFP4, 4 bit	1.32 TiB	8	16
MXFP4, AMD	1.45 TiB	8x MI355X	16

• Verified throughput, NVFP4 on 8 B300 with TP8 and MTP-3 speculative decoding: 304 output tokens per second for a single user. FP8 on TP16 reaches 307. MTP-3 is worth roughly 2.3x on the per user rate.
• RTX 5090 carries 32 GB of VRAM. NVIDIA RTX 5090 specifications
• Raw 4 bit parameter estimate: 2.4T x 0.5 bytes is 1.2 TB, which is 37.5 RTX 5090s of VRAM before any overhead, so roughly 38. Arithmetic, not a vendor figure, and it deliberately ignores KV cache, activations and the fact that no such rig exists.

The open weight release

• Announced 3 August 2026 with the open weights promised for the following week; the Qwen post says “Next week, the open weights of Qwen3.8-Max will be released, and Qwen3.8-27B is also going open-weights”. Qwen on X
• The weights shipped 12 August 2026 as the ungated Qwen/Qwen3.8-2.4T-A95B repository, with FP8 and NVFP4 checkpoints alongside. vLLM day 0 support, LMSYS SGLang day 0
• The licence is the custom Qwen3.8-Max License, not Apache 2.0. It grants use, copying, modification, publication, distribution, sublicensing, sale, deployment, hosting, fine tuning and derivative works. Products above 100 million monthly active users or $20 million monthly revenue must show the model name prominently in the interface, and a Model as a Service or AI work assistant business above $50 million of revenue in any 12 months needs a separate licence. Internal use is exempt. Qwen3.8-Max License, Forkast analysis

Not checked

Listed here because the narration asserts them and no primary source was found.

• “CoWorkBench rises by more than ten points.” Qwen3.8-Max’s CoWorkBench figure is published (76.1 on the 0902 update, per DataCamp) but no Qwen3.7-Max CoWorkBench score could be found to difference it against. No CoWorkBench figure is drawn on screen; the shot shows JobBench’s sourced 31.3 to 53.4 instead.
• “Some of Flash Next’s memory can sit in system RAM.” A plausible description of a 6B active MoE, and consistent with how these models are run, but no vendor statement was found. No figure is put on screen for it.
• CoWorkBench and JobBench eval versions. The 31.3 to 53.4 JobBench pair comes from the launch table; DataCamp reports 64.0 for the later 0902 checkpoint. Different checkpoints of the same model, not a contradiction, but the on screen figure is labelled to the launch comparison.
• The sixteen day run itself. Every figure in it (16 days, 265 commits, 127 pull requests, 151 issues) is Alibaba’s own report of its own demonstration. Nobody outside Alibaba has reproduced it, which is a point the video makes rather than hides.