DeepSeek V4 Flash and Qwen3.6 35B A3B occupy different positions in the 2026 open-weight model market. Developers select these models based on specific mathematical reasoning needs and inference budgets. DeepSeek V4 Flash focuses on low cost for reasoning at scale. Qwen3.6 35B A3B focuses on speed and multimodal input for agentic coding. GLM-4.6 targets tool use and long-context retrieval. These three models use Mixture-of-Experts architecture to maintain high capacity while keeping active parameters low.
DeepSeek V4 Flash technical profile
DeepSeek V4 Flash uses a Mixture-of-Experts design that activates 13B parameters per token from a 284B total parameter pool. The model uses CSA and HCA attention mechanisms on top of its backbone. It arrived on April 24, 2026, under an MIT license. This model holds a 1M token context window for long document reasoning. DeepInfra serves the model in FP4 precision. The input price is $0.10 per 1M tokens and the output price is $0.20 per 1M tokens. Cache hits on DeepInfra cost $0.02 per 1M tokens.
The model is the best choice for teams that prioritize low cost over peak speed. High-effort reasoning burns more time in the thinking phase before the model generates output. This reason causes the model to run at 23 tokens per second on DeepInfra. The time to first token is 0.93 seconds. DeepSeek V4 Flash scores 37 on the Artificial Analysis Intelligence Index. It scores 42% on SciCode for scientific coding quality. It scores 63% on AA-LCR for reasoning over long inputs. It also scores 39% on Terminal-Bench for agentic coding.
Qwen3.6 35B A3B technical profile
Qwen3.6 35B A3B activates 3B parameters per token from its 35B total parameter count. It routes across 256 experts with 8 routed experts and 1 shared expert per token. The model arrived in April 2026 under an Apache 2.0 license. It accepts text, image, and video input. The native context window is 262k tokens, but it reaches 1M tokens with YaRN. DeepInfra serves the model in FP8 precision. The input price is $0.15 and the output price is $0.95 per 1M tokens.
The model provides high throughput for multimodal agents. It generates at 121 tokens per second on DeepInfra. The time to first token is 0.7s. This speed is more than 5x the speed of DeepSeek V4 Flash. Qwen3.6 35B A3B scores 33 on the Artificial Analysis Intelligence Index. It scores 36% on SciCode. The model scores 64% on AA-LCR. It also scores 35% on Terminal-Bench. Qwen3.6 35B A3B is the best option for developers who need a single model for vision and code.
GLM-4.6 technical profile
Z AI built GLM-4.6 for agents that call tools and read long inputs. The model uses a Mixture-of-Experts design tuned for tool calling and retrieval. It has a 200k token context window. It launched in September 2025 as an open-weight release under the MIT license. DeepInfra serves the model in FP4. The input price is $0.43 and the output price is $1.74 per 1M tokens. Cache hits cost $0.08 per 1M tokens.
The model runs at 39 tokens per second on DeepInfra. The time to first token is 1.0s. GLM-4.6 scores 25 on the Artificial Analysis Intelligence Index. It scores 38% on SciCode. It scores 54% on AA-LCR. It scores 25% on Terminal-Bench. The value of this model depends on the difficulty of the tool use and retrieval tasks. GLM-4.6 improves on GLM-4.5 across coding, reasoning, and tool use. Does the 200k token window of GLM-4.6 provide enough space for complex retrieval tasks compared to the 1M token windows of its rivals?
Mathematical reasoning benchmark results
The HMMT February 2026 results show the performance of various models on math problems. Qwen3.7 Max scored 97.1% on this benchmark. DeepSeek V4 Pro scored 95.2%. Kimi K2.6 Thinking scored 92.7%. GLM-5.2 scored 92.5%. DeepSeek V4 Flash 0731 scored 79.9% in an unranked reasoning category. Qwen3.8 Max scored 87.6% in an unranked reasoning category.
The GSM8K benchmark measures mathematical and step-by-step reasoning through grade school word problems. A score above 85.0% indicates superior mathematical reasoning. A score between 70.0% and 85.0% indicates strong general reasoning. DeepSeek-R1-Distill-Qwen-32B in AWQ-INT4 precision achieved a score of 86.2% on GSM8K. The Qwen3.8-27B model scored 70.36% on GSM8K. The Qwen3.8-27B-FP8 model scored 68.76% on GSM8K. These numbers show that distilled reasoning models often lead in math accuracy.
| Model Name | Total Parameters | Active Parameters | Context Window | License | Input Price | Output Price |
|---|---|---|---|---|---|---|
| DeepSeek V4 Flash | 284B | 13B | 1M tokens | MIT | $0.10 | $0.20 |
| Qwen3.6 35B A3B | 35B | 3B | 262k/1M | Apache 2.0 | $0.15 | $0.95 |
| GLM-4.6 | 357B | 32B | 200k tokens | MIT | $0.43 | $1.74 |
Coding and agent performance benchmarks
Coding quality varies between models depending on the specific task requirements. DeepSeek V4 Flash leads on raw coding quality with a 42% score on SciCode. GLM-4.6 follows with a 38% score on SciCode. Qwen3.6 35B A3B scores 36% on SciCode. Qwen3.6 35B A3B leads on reasoning tasks across large codebases with a 64% score on AA-LCR. DeepSeek V4 Flash follows with a 63% score on AA-LCR. GLM-4.6 scores 54% on AA-LCR.
Agentic coding tasks involve multi-step terminal use. DeepSeek V4 Flash leads this category with a 39% score on Terminal-Bench. Qwen3.6 35B A3B follows with a 35% score on Terminal-Bench. GLM-4.6 trails with a 25% score on Terminal-Bench. The gap between DeepSeek and GLM in terminal tasks is 14 points. These results show that DeepSeek and Qwen are the stronger choices for terminal-driven agent workflows. You should evaluate these models against your specific hardware constraints before you commit to a deployment strategy.
Inference cost and speed comparisons
Inference cost for reasoning workloads depends heavily on output token prices. DeepSeek V4 Flash has an output price of $0.20 per 1M tokens. Qwen3.6 35B A3B has an output price of $0.95 per 1M tokens. GLM-4.6 has an output price of $1.74 per 1M tokens. Output prices for GLM-4.6 are almost 9x the output prices for DeepSeek V4 Flash. A reasoning model emits long chains of thought, so output price compounds across every request. DeepSeek V4 Flash is the cheapest model to run at volume.
Speed is another differentiator for production deployments. Qwen3.6 35B A3B generates at 121 tokens per second. GLM-4.6 generates at 39 tokens per second. DeepSeek V4 Flash generates at 23 tokens per second. Qwen3.6 35B A3B is more than 3x faster than GLM-4.6 and 5x faster than DeepSeek V4 Flash.
| Feature | DeepSeek V4 Flash | Qwen3.6 35B A3B | GLM-4.6 |
|---|---|---|---|
| Generation Speed | 23 tokens/s | 121 tokens/s | 39 tokens/s |
| Time to First Token | 0.93s | 0.7s | 1.0s |
| Input Price (1M) | $0.10 | $0.15 | $0.43 |
| Output Price (1M) | $0.20 | $0.95 | $1.74 |
| Cache Hit Price | $0.02 | N/A | $0.08 |
Licensing and model selection
Licensing terms dictate how businesses deploy these models in production. DeepSeek V4 Flash and GLM-4.6 use the MIT license. This license allows for commercial use without usage restrictions. Qwen3.6 35B A3B uses the Apache 2.0 license. This license allows for commercial use, modification, and redistribution without restrictions. Kimi K2.6 uses a modified MIT license. This license requires users with 100M monthly active users or $20M monthly revenue to display the model name in the product UI.
Selection depends on the primary workload of the application. DeepSeek V4 Flash is the choice for cost-sensitive reasoning at scale. It provides the lowest inference cost and a large context window. Qwen3.6 35B A3B is the choice for speed and multimodal input. It provides the highest generation speed and handles image and video input. GLM-4.6 is the choice for tool use and retrieval. It targets multi-step agent loops and large retrieval contexts. DeepSeek V4 Flash leads on cost-effective reasoning, while Qwen3.6 35B A3B leads on speed and multimodal tasks.
