Qwen3.8-Flash-Next (180B parameters) compressed to run on one 16 GB NVIDIA graphics card with 64 GB of RAM, with Glyd's own engine. Hummingbird is Glyd's most compact level.
Run it: github.com/surya-koritala/glyd-qwen-3.8-flash
A space shooter written in one prompt on an RTX 4080 SUPER, unedited (precise mode, default settings).
[!Note] These files are in Glyd's own format and run only with the Glyd engine, which
./setup.shin the GitHub repository installs. They do not load in Transformers, vLLM or llama.cpp. For those, use Qwen's original weights.
Linux on x86-64, an NVIDIA RTX 30, 40 or 50 series card (or newer) with 16 GB or more and driver 580 or newer, 64 GB of RAM, about 90 GB of free disk:
git clone https://github.com/surya-koritala/glyd-qwen-3.8-flash && cd glyd-qwen-3.8-flash
./setup.sh
setup.sh installs the engine (no sudo, nothing else to install), downloads model/ and the mode you choose from this repository (it
resumes if it stops), and opens a chat app at http://127.0.0.1:8080. Apps and coding agents connect to the OpenAI-compatible API at
http://127.0.0.1:8080/v1 or the Anthropic-compatible one at http://127.0.0.1:8080/v1/messages.
| Folder | What it holds | Size |
|---|---|---|
model/ | Qwen's tokenizer, chat template and configs, and the weights outside the experts (attention, embeddings, n-gram table) | 37 GB |
precise/ | the experts for precise mode (the default) | 45 GB |
pack/ | the experts for fast mode | 31 GB |
You need model/ and one of the two expert folders: 82 GB for precise mode, 68 GB for fast mode.
The experts (120.8B of the parameters) are compressed according to use: the experts the model routes to most keep the most detail. The GPU holds the busiest experts, RAM holds all of them, and the CPU computes the rest while the GPU works.
On one RTX 4080 SUPER (16 GB), a Ryzen 9 7950X3D and 64 GB of RAM:
| Precise | Fast | |
|---|---|---|
| Writing, in a chat (the default settings) | 117 tokens/s | 135 tokens/s |
| Writing, top speed (a short prompt, greedy) | 145 tokens/s | 169-173 tokens/s |
| Reading a 32,000-token prompt | about 7 s | about 7 s |
| Reading a 261,577-token prompt | 66 s |
Contexts up to 262,144 tokens (the model's native length). Follow-up messages reuse the conversation, so they start in under a second.
Inspect AI's tasks as they ship, 150 questions each, greedy, thinking off, a 3,072-token cap:
| Precise | Fast | |
|---|---|---|
| GSM8K | 148 / 150 | 148 / 150 |
| MMLU-Pro | 124 / 150 | 118 / 150 |
| A full game in one prompt (Star Drift) | 3 of 6 playable | 1 of 6 |
The logs and the scripts that produced them are in the GitHub repository (results/, bench/).
Limits of this release: text only (the base model's vision encoder is not supported yet), NVIDIA GPUs with 16 GB or more on Linux, and one request at a time.
Qwen3.8-Flash-Next by the Qwen team: 125B parameters in a mixture of experts with 6B active per token, plus a 51B n-gram embedding and a 4B multi-token prediction layer. It has 48 layers, each with 512 routed experts, of which 10 (plus 1 shared) run for each token, and it pairs Gated DeltaNet with Qwen Sparse Attention for long contexts. Details: the original model card.
These weights are derived from Qwen3.8-Flash-Next and are under its license, the Qwen Community License 1.0 (LICENSE). The Glyd engine that runs them is released separately, in compiled form, under the Business Source License 1.1: free for personal, research and other non-commercial use (see the GitHub repository).
Thanks to the Qwen team for releasing Qwen3.8-Flash-Next with open weights.
Qwen3.8-Flash-Next (180B parameters) compressed to run on one 16 GB NVIDIA graphics card with 64 GB of RAM, with Glyd's own engine. Hummingbird is Glyd's most compact level.
Run it: github.com/surya-koritala/glyd-qwen-3.8-flash
A space shooter written in one prompt on an RTX 4080 SUPER, unedited (precise mode, default settings).
[!Note] These files are in Glyd's own format and run only with the Glyd engine, which
./setup.shin the GitHub repository installs. They do not load in Transformers, vLLM or llama.cpp. For those, use Qwen's original weights.
Linux on x86-64, an NVIDIA RTX 30, 40 or 50 series card (or newer) with 16 GB or more and driver 580 or newer, 64 GB of RAM, about 90 GB of free disk:
git clone https://github.com/surya-koritala/glyd-qwen-3.8-flash && cd glyd-qwen-3.8-flash
./setup.sh
setup.sh installs the engine (no sudo, nothing else to install), downloads model/ and the mode you choose from this repository (it
resumes if it stops), and opens a chat app at http://127.0.0.1:8080. Apps and coding agents connect to the OpenAI-compatible API at
http://127.0.0.1:8080/v1 or the Anthropic-compatible one at http://127.0.0.1:8080/v1/messages.
| Folder | What it holds | Size |
|---|---|---|
model/ | Qwen's tokenizer, chat template and configs, and the weights outside the experts (attention, embeddings, n-gram table) | 37 GB |
precise/ | the experts for precise mode (the default) | 45 GB |
pack/ | the experts for fast mode | 31 GB |
You need model/ and one of the two expert folders: 82 GB for precise mode, 68 GB for fast mode.
The experts (120.8B of the parameters) are compressed according to use: the experts the model routes to most keep the most detail. The GPU holds the busiest experts, RAM holds all of them, and the CPU computes the rest while the GPU works.
On one RTX 4080 SUPER (16 GB), a Ryzen 9 7950X3D and 64 GB of RAM:
| Precise | Fast | |
|---|---|---|
| Writing, in a chat (the default settings) | 117 tokens/s | 135 tokens/s |
| Writing, top speed (a short prompt, greedy) | 145 tokens/s | 169-173 tokens/s |
| Reading a 32,000-token prompt | about 7 s | about 7 s |
| Reading a 261,577-token prompt | 66 s |
Contexts up to 262,144 tokens (the model's native length). Follow-up messages reuse the conversation, so they start in under a second.
Inspect AI's tasks as they ship, 150 questions each, greedy, thinking off, a 3,072-token cap:
| Precise | Fast | |
|---|---|---|
| GSM8K | 148 / 150 | 148 / 150 |
| MMLU-Pro | 124 / 150 | 118 / 150 |
| A full game in one prompt (Star Drift) | 3 of 6 playable | 1 of 6 |
The logs and the scripts that produced them are in the GitHub repository (results/, bench/).
Limits of this release: text only (the base model's vision encoder is not supported yet), NVIDIA GPUs with 16 GB or more on Linux, and one request at a time.
Qwen3.8-Flash-Next by the Qwen team: 125B parameters in a mixture of experts with 6B active per token, plus a 51B n-gram embedding and a 4B multi-token prediction layer. It has 48 layers, each with 512 routed experts, of which 10 (plus 1 shared) run for each token, and it pairs Gated DeltaNet with Qwen Sparse Attention for long contexts. Details: the original model card.
These weights are derived from Qwen3.8-Flash-Next and are under its license, the Qwen Community License 1.0 (LICENSE). The Glyd engine that runs them is released separately, in compiled form, under the Business Source License 1.1: free for personal, research and other non-commercial use (see the GitHub repository).
Thanks to the Qwen team for releasing Qwen3.8-Flash-Next with open weights.