Ornith-1.0-9B — ASHQ1 Quantization (MTP-Aware)
ASHQ1 quantization of Ornith-1.0-9B with explicit MTP (Multi-Token Prediction) handling — MTP heads are kept at Q8_0 and excluded from the classifier budget, preserving their quality while the priority queue optimises the main transformer layers.
Huge thanks to protoLabsAI/Ornith-1.0-9B-MTP-GGUF for providing the BF16 source with MTP heads — we adapted it with ASHQ1 to deliver both speed and quality.
Note: File names contain "BF16" for HuggingFace parser compatibility — these are ASHQ1 quants, not BF16.
Quants
| File | Size | Method | Description |
|---|---|---|---|
Ornith-1.0-9B-MTP-BF16-ASHQ1-4850.gguf |
4.7 GB | ASHQ1 (v6, --allow-q3-or-lower) |
Tiny — uses IQ tiers to fit 4.7 GB, MTP heads at Q8_0 |
Ornith-1.0-9B-MTP-BF16-ASHQ1-5500.gguf |
5.4 GB | ASHQ1 (v6) | Compact — beats i1-Q6_K (7.0 GB), 23% smaller |
Ornith-1.0-9B-MTP-BF16-ASHQ1-6500.gguf |
6.4 GB | ASHQ1 (v6) | Balanced — near i1-Q8_0 quality at 33% smaller |
Ornith-1.0-9B-MTP-BF16-ASHQ1-7300.gguf |
7.2 GB | ASHQ1 (v6) | Best — surpasses i1-Q8_0 quality, 25% smaller |
All comparisons are against i1 quants (uniform quantization with bartowski's Ornith imatrix). ASHQ1 consistently beats them at smaller sizes: the priority queue allocates bits where they matter most. See wepiqx/ASHQ1 for full benchmarks across architectures. MTP heads are deployed at Q8_0 and excluded from the classifier budget.
Note: The 4850 variant uses IQ tiers via
--allow-q3-or-lowerto fit the tight budget — quality is roughly on par with Q4_K_S, while 5500+ delivers a significant step up. It's best suited for lighter tasks, while coding and agentic workloads may prefer 5500+. Feel free to try it either way.
How ASHQ1 Works
See wepiqx/ASHQ1 for the full method, source code, and benchmarks.
Usage
llama.cpp
# Chat mode with built-in Jinja template (recommended)
llama-cli \
-m Ornith-1.0-9B-MTP-BF16-ASHQ1-6500.gguf \
--jinja \
-ngl 99 \
-c 8192
Recommended sampling: temperature 0.6–1.0, top_p 0.95, top_k 20.
Ollama
The model's built-in Jinja template is auto-detected by Ollama — no need to specify TEMPLATE manually.
- Create a
Modelfile:
FROM ./Ornith-1.0-9B-MTP-BF16-ASHQ1-6500.gguf
PARAMETER num_ctx 8192
PARAMETER temperature 0.6
PARAMETER top_k 20
PARAMETER top_p 0.95
- Build and run:
ollama create ornith-ashq1-mtp-6500 -f Modelfile
ollama run ornith-ashq1-mtp-6500
LM Studio
- Open LM Studio
- Drag the desired GGUF file (4850, 5500, 6500, or 7300) into the app
- Set GPU Offload to 99 layers
- Set context length to 8192+
- Set preset sampling:
temperature 0.6,top_p 0.95,top_k 20 - Start chatting
Quantization Configs
Generated by ASHQ1 priority queue (v6) — files in this repo were produced with these exact llama-quantize arguments.
4850 Config (--allow-q3-or-lower)
--output-tensor-type Q5_K
--token-embedding-type Q5_K
--tensor-type "(blk|BLK)\.(32)\.nextn=Q8_0"
--tensor-type "(blk|BLK)\.(32)\.ffn_down=Q4_K"
--tensor-type "(blk|BLK)\.(31)\.ffn_down=IQ4_XS"
--tensor-type "(blk|BLK)\.(30)\.ffn_down=IQ3_S"
--tensor-type "(blk|BLK)\.(32)\.attn_v=Q4_K"
--tensor-type "(blk|BLK)\.(32)\.attn_q=Q4_K"
--tensor-type "(blk|BLK)\.(32)\.attn_k=Q4_K"
--tensor-type "(blk|BLK)\.((?:26|27|28|29))\.ffn_down=IQ3_XXS"
--tensor-type "(blk|BLK)\.((?:22|23|24|25))\.ffn_down=IQ2_S"
--tensor-type "(blk|BLK)\.((?:0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21))\.ffn_down=IQ2_XXS"
--tensor-type "(blk|BLK)\.(3|7|11|15|19|23|27|(?:31|32))\.attn_k_norm=F16"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_norm=F16"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_a=F16"
--tensor-type "(blk|BLK)\.((?:0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32))\.attn_norm=F16"
--tensor-type "(blk|BLK)\.(3|7|11|15|19|23|27|(?:31|32))\.attn_q_norm=F16"
--tensor-type "(blk|BLK)\.((?:0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32))\.post_attention_norm=F16"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_conv1d=F16"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_dt=F16"
--tensor-type "(blk|BLK)\.(3|7|11|15|19|23|27|31)\.attn_q=Q5_K"
--tensor-type "(blk|BLK)\.(0|6|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.attn_qkv=Q5_K"
--tensor-type "(blk|BLK)\.(0|6|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.attn_gate=Q5_K"
--tensor-type "(blk|BLK)\.(0|6|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_alpha=Q5_K"
--tensor-type "(blk|BLK)\.(0|6|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_beta=Q5_K"
--tensor-type "(blk|BLK)\.(3|7|11|15|19|23|27|31)\.attn_k=Q5_K"
--tensor-type "(blk|BLK)\.(3|7|11|15|19|23|27|31)\.attn_v=Q5_K"
--tensor-type "(blk|BLK)\.(3|7|11|15|19|23|27|(?:31|32))\.attn_output=Q4_K"
--tensor-type "(blk|BLK)\.((?:0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32))\.ffn_gate=Q4_K"
--tensor-type "(blk|BLK)\.((?:0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32))\.ffn_up=Q4_K"
--tensor-type "(blk|BLK)\.([1-2]|[4-5])\.attn_gate=Q4_K"
--tensor-type "(blk|BLK)\.([1-2]|[4-5])\.ssm_beta=Q4_K"
--tensor-type "(blk|BLK)\.([1-2]|[4-5])\.ssm_alpha=Q4_K"
--tensor-type "(blk|BLK)\.([1-2]|[4-5])\.attn_qkv=Q4_K"
--tensor-type "(blk|BLK)\.(28|30)\.ssm_out=IQ4_XS"
--tensor-type "(blk|BLK)\.((?:24|25)|29)\.ssm_out=IQ3_S"
--tensor-type "(blk|BLK)\.((?:21|22)|26)\.ssm_out=IQ3_XXS"
--tensor-type "(blk|BLK)\.(18|20)\.ssm_out=IQ2_S"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17))\.ssm_out=IQ2_XXS"
--tensor-type ".*output_norm.*=F16"
5500 Config (base=Q5_K_M)
--output-tensor-type Q5_K
--token-embedding-type Q5_K
--tensor-type "(blk|BLK)\.(32)\.nextn=Q8_0"
--tensor-type "(blk|BLK)\.(3)\.attn_q=Q6_K"
--tensor-type "(blk|BLK)\.(3)\.attn_k=Q6_K"
--tensor-type "(blk|BLK)\.(31)\.ffn_down=Q5_K"
--tensor-type "(blk|BLK)\.((?:20|21)|(?:24|25)|28|30)\.attn_qkv=Q6_K"
--tensor-type "(blk|BLK)\.(3|32)\.attn_v=Q6_K"
--tensor-type "(blk|BLK)\.((?:20|21)|(?:24|25)|28|30)\.ssm_beta=Q6_K"
--tensor-type "(blk|BLK)\.((?:20|21)|(?:24|25)|28|30)\.ssm_alpha=Q6_K"
--tensor-type "(blk|BLK)\.((?:20|21)|(?:24|25)|28|30)\.attn_gate=Q6_K"
--tensor-type "(blk|BLK)\.(31)\.attn_output=Q5_K"
--tensor-type "(blk|BLK)\.(30)\.ssm_out=Q5_K"
--tensor-type "(blk|BLK)\.(1)\.attn_gate=Q4_K"
--tensor-type "(blk|BLK)\.(1)\.ssm_alpha=Q4_K"
--tensor-type "(blk|BLK)\.(32)\.attn_q=Q4_K"
--tensor-type "(blk|BLK)\.(1)\.ssm_beta=Q4_K"
--tensor-type "(blk|BLK)\.(1)\.attn_qkv=Q4_K"
--tensor-type "(blk|BLK)\.((?:0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30)|32)\.ffn_down=Q4_K"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_conv1d=F16"
--tensor-type "(blk|BLK)\.((?:0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32))\.post_attention_norm=F16"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_norm=F16"
--tensor-type "(blk|BLK)\.((?:0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32))\.attn_norm=F16"
--tensor-type "(blk|BLK)\.(3|7|11|15|19|23|27|(?:31|32))\.attn_k_norm=F16"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_dt=F16"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_a=F16"
--tensor-type "(blk|BLK)\.(3|7|11|15|19|23|27|(?:31|32))\.attn_q_norm=F16"
--tensor-type "(blk|BLK)\.((?:26|27|28|29|30|31))\.ffn_up=Q5_K"
--tensor-type "(blk|BLK)\.(0|2|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|22|26|29)\.attn_qkv=Q5_K"
--tensor-type "(blk|BLK)\.(0|2|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|22|26|29)\.ssm_alpha=Q5_K"
--tensor-type "(blk|BLK)\.(7|11|15|19|23|27|(?:31|32))\.attn_k=Q5_K"
--tensor-type "(blk|BLK)\.(7|11|15|19|23|27|31)\.attn_v=Q5_K"
--tensor-type "(blk|BLK)\.(0|2|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|22|26|29)\.attn_gate=Q5_K"
--tensor-type "(blk|BLK)\.(7|11|15|19|23|27|31)\.attn_q=Q5_K"
--tensor-type "(blk|BLK)\.(0|2|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|22|26|29)\.ssm_beta=Q5_K"
--tensor-type "(blk|BLK)\.((?:26|27|28|29|30|31))\.ffn_gate=Q5_K"
--tensor-type "(blk|BLK)\.((?:0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25)|32)\.ffn_up=Q4_K"
--tensor-type "(blk|BLK)\.((?:0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25)|32)\.ffn_gate=Q4_K"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29))\.ssm_out=Q4_K"
--tensor-type "(blk|BLK)\.(3|7|11|15|19|23|27|32)\.attn_output=Q4_K"
--tensor-type ".*output_norm.*=F16"
6500 Config (base=Q5_K_M)
--output-tensor-type Q5_K
--token-embedding-type Q5_K
--tensor-type "(blk|BLK)\.(32)\.nextn=Q8_0"
--tensor-type "(blk|BLK)\.(31)\.ffn_down=Q6_K"
--tensor-type "(blk|BLK)\.(0|(?:9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_beta=Q8_0"
--tensor-type "(blk|BLK)\.(0|(?:9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.attn_gate=Q8_0"
--tensor-type "(blk|BLK)\.(3|7|15|19|23|27|31)\.attn_v=Q8_0"
--tensor-type "(blk|BLK)\.(0|(?:9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.attn_qkv=Q8_0"
--tensor-type "(blk|BLK)\.(0|(?:9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_alpha=Q8_0"
--tensor-type "(blk|BLK)\.(3|7|15|19|23|27|31)\.attn_k=Q8_0"
--tensor-type "(blk|BLK)\.(3|7|15|19|23|27|31)\.attn_q=Q8_0"
--tensor-type "(blk|BLK)\.(11)\.attn_q=Q6_K"
--tensor-type "(blk|BLK)\.(11)\.attn_k=Q6_K"
--tensor-type "(blk|BLK)\.(11)\.attn_v=Q6_K"
--tensor-type "(blk|BLK)\.([1-2]|[4-6]|8)\.ssm_beta=Q6_K"
--tensor-type "(blk|BLK)\.([1-2]|[4-6]|8)\.attn_qkv=Q6_K"
--tensor-type "(blk|BLK)\.((?:14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31))\.ffn_gate=Q6_K"
--tensor-type "(blk|BLK)\.([1-2]|[4-6]|8)\.ssm_alpha=Q6_K"
--tensor-type "(blk|BLK)\.((?:28|29|30))\.ssm_out=Q6_K"
--tensor-type "(blk|BLK)\.((?:14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31))\.ffn_up=Q6_K"
--tensor-type "(blk|BLK)\.([1-2]|[4-6]|8)\.attn_gate=Q6_K"
--tensor-type "(blk|BLK)\.(27|31)\.attn_output=Q6_K"
--tensor-type "(blk|BLK)\.(32)\.attn_v=Q5_K"
--tensor-type "(blk|BLK)\.(32)\.attn_k=Q4_K"
--tensor-type "(blk|BLK)\.(32)\.attn_q=Q4_K"
--tensor-type "(blk|BLK)\.((?:26|27|28|29|30))\.ffn_down=Q5_K"
--tensor-type "(blk|BLK)\.((?:0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25)|32)\.ffn_down=Q4_K"
--tensor-type "(blk|BLK)\.(3|7|11|15|19|23|27|(?:31|32))\.attn_q_norm=F16"
--tensor-type "(blk|BLK)\.((?:0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32))\.post_attention_norm=F16"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_norm=F16"
--tensor-type "(blk|BLK)\.((?:0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32))\.attn_norm=F16"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_a=F16"
--tensor-type "(blk|BLK)\.(3|7|11|15|19|23|27|(?:31|32))\.attn_k_norm=F16"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_conv1d=F16"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_dt=F16"
--tensor-type "(blk|BLK)\.(18|(?:20|21|22)|(?:24|25|26))\.ssm_out=Q5_K"
--tensor-type "(blk|BLK)\.((?:4|5|6|7|8|9|10|11|12|13))\.ffn_gate=Q5_K"
--tensor-type "(blk|BLK)\.((?:4|5|6|7|8|9|10|11|12|13))\.ffn_up=Q5_K"
--tensor-type "(blk|BLK)\.(19|23)\.attn_output=Q5_K"
--tensor-type "(blk|BLK)\.([0-3]|32)\.ffn_up=Q4_K"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17))\.ssm_out=Q4_K"
--tensor-type "(blk|BLK)\.([0-3]|32)\.ffn_gate=Q4_K"
--tensor-type "(blk|BLK)\.(3|7|11|15|32)\.attn_output=Q4_K"
--tensor-type ".*output_norm.*=F16"
7300 Config (base=Q5_K_M)
--output-tensor-type Q5_K
--token-embedding-type Q5_K
--tensor-type "(blk|BLK)\.(32)\.nextn=Q8_0"
--tensor-type "(blk|BLK)\.(31)\.ffn_down=Q8_0"
--tensor-type "(blk|BLK)\.(27|31)\.attn_output=Q8_0"
--tensor-type "(blk|BLK)\.(3|7|11|15|19|23|27|31)\.attn_v=Q8_0"
--tensor-type "(blk|BLK)\.(28|30)\.ssm_out=Q8_0"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.attn_qkv=Q8_0"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.attn_gate=Q8_0"
--tensor-type "(blk|BLK)\.((?:16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31))\.ffn_up=Q8_0"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_beta=Q8_0"
--tensor-type "(blk|BLK)\.((?:16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31))\.ffn_gate=Q8_0"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_alpha=Q8_0"
--tensor-type "(blk|BLK)\.(3|7|11|15|19|23|27|31)\.attn_q=Q8_0"
--tensor-type "(blk|BLK)\.(3|7|11|15|19|23|27|31)\.attn_k=Q8_0"
--tensor-type "(blk|BLK)\.(32)\.attn_k=Q6_K"
--tensor-type "(blk|BLK)\.((?:22|23|24|25|26|27|28|29|30))\.ffn_down=Q6_K"
--tensor-type "(blk|BLK)\.((?:4|5|6|7|8|9|10|11|12|13|14|15))\.ffn_up=Q6_K"
--tensor-type "(blk|BLK)\.((?:4|5|6|7|8|9|10|11|12|13|14|15))\.ffn_gate=Q6_K"
--tensor-type "(blk|BLK)\.(18|(?:20|21|22)|(?:24|25|26)|29)\.ssm_out=Q6_K"
--tensor-type "(blk|BLK)\.(19|23)\.attn_output=Q6_K"
--tensor-type "(blk|BLK)\.(32)\.attn_v=Q5_K"
--tensor-type "(blk|BLK)\.(32)\.ffn_gate=Q4_K"
--tensor-type "(blk|BLK)\.(32)\.ffn_up=Q4_K"
--tensor-type "(blk|BLK)\.(32)\.attn_q=Q4_K"
--tensor-type "(blk|BLK)\.((?:18|19|20|21))\.ffn_down=Q5_K"
--tensor-type "(blk|BLK)\.((?:0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17)|32)\.ffn_down=Q4_K"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_dt=F16"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_norm=F16"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_a=F16"
--tensor-type "(blk|BLK)\.((?:0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32))\.attn_norm=F16"
--tensor-type "(blk|BLK)\.((?:0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32))\.post_attention_norm=F16"
--tensor-type "(blk|BLK)\.(3|7|11|15|19|23|27|(?:31|32))\.attn_k_norm=F16"
--tensor-type "(blk|BLK)\.(3|7|11|15|19|23|27|(?:31|32))\.attn_q_norm=F16"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|(?:8|9|10)|(?:12|13|14)|(?:16|17|18)|(?:20|21|22)|(?:24|25|26)|(?:28|29|30))\.ssm_conv1d=F16"
--tensor-type "(blk|BLK)\.([0-3])\.ffn_gate=Q5_K"
--tensor-type "(blk|BLK)\.([0-3])\.ffn_up=Q5_K"
--tensor-type "(blk|BLK)\.((?:9|10)|(?:12|13|14)|(?:16|17))\.ssm_out=Q5_K"
--tensor-type "(blk|BLK)\.(7|11|15)\.attn_output=Q5_K"
--tensor-type "(blk|BLK)\.([0-2]|[4-6]|8)\.ssm_out=Q4_K"
--tensor-type "(blk|BLK)\.(3|32)\.attn_output=Q4_K"
--tensor-type ".*output_norm.*=F16"
Hardware Requirements
| Variant | Size (GGUF) | VRAM (ctx 8192) | Minimum GPU |
|---|---|---|---|
| 4850 | 4.7 GB | ~5.8 GB | 8 GB |
| 5500 | 5.4 GB | ~6.5 GB | 8 GB |
| 6500 | 6.4 GB | ~7.5 GB | 8 GB |
| 7300 | 7.2 GB | ~8.3 GB | 12 GB |
Notes
- Source BF16: protoLabsAI/Ornith-1.0-9B-MTP-GGUF — huge thanks to protoLabsAI for the MTP-ready BF16
- Imatrix: bartowski's Ornith imatrix (248 entries, 6 datasets)
- Architecture: Qwen3.5-9B (32 layers, GQA, SSM, MTP, 262k context)
- MTP heads deployed at Q8_0
- License: MIT
- Quantization and tuning by wepiqx
- Built with llama.cpp