Ledger on GitHub. Hub here. Branches elsewhere.
Qwen 3.8 Max at $2 per million tokens
Qwen 3.8 Max at $2 per million tokens
A 2.4-trillion-parameter mixture-of-experts model from Alibaba reached API pricing of $2 input and $6 output per million tokens, with 95 billion active parameters and a 1-million-token context window.
Qwen 3.8 Max was confirmed at that $2/$6 rate across three separate URLs. The Alibaba model page indexed it as 2.4T total parameters, 95B active, and 1M context. Two other sources repeated the same pricing and context figures without variance.
GLM-5.3-Flash arrived as a 320-billion-parameter total, 18-billion-active multimodal mixture-of-experts under the MIT license. Context extends to 1 million tokens. API pricing was reported at $0.15 input and $0.50 output per million tokens. The MIT license and parameter counts appeared on both the original announcement and a model-repository mirror.
DeepSeek V4-Pro was indexed as MIT-licensed open weights with 1.65 trillion parameters, 61 layers, 384 experts, and 6 active experts per token. Context runs to 1 million tokens. API cost was listed at $0.435 input and $0.87 output per million tokens. A separate source described DeepSeek V4-Flash-Vision-Exp as matching V4-Flash on text and improving multimodal agent benchmarks, without parameter counts.
Moonshot Kimi K3 was described as a 2.8-trillion-parameter open-weight model with downloadable weights, though the table row ends mid-sentence and no pricing appeared in the harvest.
Benchmark figures for Qwen 3.8 Max included a PaperBench score of 93.0, Terminal Bench 2.1 at 86.6, and SWE-bench Pro at 67.7. Those numbers carried a sensitivity flag in the table because benchmark reproducibility and task definitions vary. No independent verification URLs were included.
The pricing gap between Qwen 3.8 Max at $2 input and GLM-5.3-Flash at $0.15 input spans more than a 13× ratio. Qwen 3.8 Max activates 95 billion parameters per token while GLM-5.3-Flash activates 18 billion. Context and API availability match at 1 million tokens for both. DeepSeek V4-Pro slots between them at $0.435 input, activating 6 experts from a pool of 384 across 1.65 trillion total parameters.
All three models carry open-weight or MIT licenses, meaning practitioners can compare local inference cost to API cost without vendor lock. The 1-million-token context window is consistent across the set, removing that variable from the decision tree.
No hardware requirements, throughput numbers, or batch-size limits appeared in the table. No information about embedding endpoints, vision modality pricing differences, or rate limits was harvested.
| Model | Total params | Active params | Context | Input $/1M | Output $/1M | License | URL |
|---|---|---|---|---|---|---|---|
| Qwen 3.8 Max | 2.4T | 95B | 1M | $2.00 | $6.00 | not in sources | https://www.qwencloud.com/models/qwen3.8-2.4t-a95b |
| GLM-5.3-Flash | 320B | 18B | 1M | $0.15 | $0.50 | MIT | https://lmstudio.ai/models/glm-5.3-flash |
| DeepSeek V4-Pro | 1.65T | 384 experts, 6 active | 1M | $0.435 | $0.87 | MIT | https://wujec.ai/en/news/deepseek-v4-pro-0813-official-release |
Three lines for practitioners follow.
Act: Download GLM-5.3-Flash weights under MIT license and measure local inference cost against the $0.15/$0.50 API rate for your context distribution, then repeat for DeepSeek V4-Pro at $0.435/$0.87 to find the break-even volume.
Watch: Qwen 3.8 Max benchmark claims on PaperBench 93.0 and SWE-bench Pro 67.7 need independent reproduction runs with public task definitions and prompt templates before you route production traffic based on those numbers.
Ignore: Any inference provider that does not publish per-million-token pricing, active parameter counts, and context-window limits in a single table—variance across three sources for Qwen and two for GLM showed zero drift, so anything less transparent is a pricing negotiation surface you do not want.