Records, runnable research and changes that leave a machine out. tag:computecurve.com,2026:feed/all 2026-09-09T00:00:00Z radar tag:computecurve.com,2026:2026-09-09/record/Qwen3-8B/strix-halo/1/45.4 2026-09-09T00:00:00Z

Beats the previous published best of 34.0 for the same model on the same setup, by 34%. Recipe: lasshow/strix-halo-llm-lab. Stack: llama.cpp, UD-IQ4_XS, Q4.

tag:computecurve.com,2026:2026-09-09/record/Qwen3-30B-A3B/strix-halo/1/88.4 2026-09-09T00:00:00Z

Beats the previous published best of 86.1 for the same model on the same setup, by 3%. Recipe: tahircengiz/strix-halo-llm-benchmarks. Stack: llama.cpp, Q4_K_M.

tag:computecurve.com,2026:2026-09-08/excluye/TensorRT-LLM/18446 2026-09-08T00:00:00Z

Leaves out GeForce RTX 5090, RTX PRO 6000 Blackwell, NVIDIA DGX Spark (GB10). On sm_100/103 enable enable_heuristic_topk and use_self_sampling_topk; sm_120/121 restarts get a radix fallback with no GVR speedup. Cost: restart.

tag:computecurve.com,2026:2026-09-07/excluye/sglang/38293 2026-09-07T00:00:00Z

Leaves out AMD Instinct MI300X, AMD Ryzen AI Max+ 395 (Strix Halo). On ROCm and Ascend NPU, --enable-prefill-cp now fails at startup, so remove it or set cp_size=1 before restarting. Cost: restart.

tag:computecurve.com,2026:2026-09-04/record/GLM-5.3-Flash/strix-halo/1/17.37 2026-09-04T00:00:00Z

Beats the previous published best of 14.63 for the same model on the same setup, by 19%. Recipe: YanissAmz/strix-halo-llm-serving. Stack: llama.cpp, UD-IQ3_XXS.

tag:computecurve.com,2026:2026-09-03/record/GLM-5.3-Flash/dgx-spark/3/93.0 2026-09-03T00:00:00Z

Beats the previous published best of 69.0 for the same model on the same setup, by 35%. Recipe: NNNtrance/GLM-5.3-Flash-NVFP4-TP3-3x-DGX-Spark. Stack: vllm, NVFP4.

tag:computecurve.com,2026:2026-09-03/paper/2609.03430 2026-09-03T00:00:00Z

A parameter-free cache policy ships code and a drop-in vLLM recipe that evicts tokens per head without scoring. It ships code and a drop-in vLLM deployment recipe that raises throughput on reasoning workloads without any new hardware. Releases: code.

tag:computecurve.com,2026:2026-09-03/paper/2609.03796 2026-09-03T00:00:00Z

Weights, training code, and detailed recipes ship for local run or fine-tune of the model with a frozen VLM backbone. The released weights and recipes let you run or fine-tune a strong open image generator locally. Releases: both.

tag:computecurve.com,2026:2026-09-03/paper/2609.04010 2026-09-03T00:00:00Z

Discrete diffusion weights let a single model generate multiple tokens in parallel with no draft model, code and checkpoints released. Uno checkpoints are released and can be run to get up to 3x generation speedups over the base AR model with no draft model needed. Releases: both.

tag:computecurve.com,2026:2026-09-03/paper/2609.04098 2026-09-03T00:00:00Z

Minima releases a W4A4 checkpoint quantizing all 496 linear layers, Gated DeltaNet included, for consumer GPUs. A ready-to-run 17.5 GiB W4A4 checkpoint of a hybrid LLM that matches BF16, useful for fitting large models on consumer GPUs. Releases: weights.

tag:computecurve.com,2026:2026-09-03/excluye/cutlass/3410 2026-09-03T00:00:00Z

Leaves out NVIDIA DGX Spark (GB10). A fresh CUDA 12.8 build no longer emits sm_121 cubins, so DGX Spark must upgrade to CUDA 12.9 or re-add 121/121a to CUTLASS_NVCC_ARCHS_SUPPORTED by hand. Cost: upgrade.

tag:computecurve.com,2026:2026-09-02/paper/2609.01956 2026-09-02T00:00:00Z

The pip-installable drop-in layer removes Cox-de Boor recursion and scatter-gather, shipping as a production-ready open-source package. It ships a pip-installable drop-in KAN layer that removes recursion and scatter-gather, so KAN inference can actually run fast on a local GPU. Releases: code.

tag:computecurve.com,2026:2026-09-02/paper/2609.03181 2026-09-02T00:00:00Z

A 3B MoE document parser ships with a drop-in speculative decoding recipe and verifiable rewards. It ships a ready-to-run OCR model that doubles decoding speed on low-budget GPUs like the L4. Releases: weights.

tag:computecurve.com,2026:2026-09-01/record/Qwen3.8-27B/strix-halo/1/94.7 2026-09-01T00:00:00Z

Beats the previous published best of 33.8 for the same model on the same setup, by 180%. Recipe: simongonzalezdc/resonant-mtp-recipe. Stack: llama.cpp, Q4_K_M, UD-Q4_K_XL.

tag:computecurve.com,2026:2026-09-01/record/DeepSeek-V4-Flash-Vision-Exp/dgx-spark/2/66.4 2026-09-01T00:00:00Z

Beats the previous published best of 26.2 for the same model on the same setup, by 153%. Recipe: shige0501/dsv4-vision-spark. Stack: vllm, FP8.

tag:computecurve.com,2026:2026-09-01/paper/2609.00791 2026-09-01T00:00:00Z

Weights, training code, data mixtures and configurations ship with the model for self-hosted inference. Fully open weights and training code are released, so a user can download and run the model on their own hardware. Releases: both.

tag:computecurve.com,2026:2026-09-01/paper/2609.01046 2026-09-01T00:00:00Z

Merged weights land on Hugging Face under Apache-2.0 for low-latency local inference. You can download and run the merged weights as a fast, low-latency guardrail on your own hardware. Releases: weights.

tag:computecurve.com,2026:2026-08-31/record/Qwen3.8-Flash-Next/dgx-spark/1/110.4 2026-08-31T00:00:00Z

Beats the previous published best of 50.6 for the same model on the same setup, by 118%. Recipe: bemlerlabs/qwen3.8-flash-next-dgx-spark-sglang. Stack: sglang, NVFP4.

tag:computecurve.com,2026:2026-08-31/record/GLM-5.3-Flash/strix-halo/1/14.63 2026-08-31T00:00:00Z

Beats the previous published best of 9.31 for the same model on the same setup, by 57%. Recipe: tomasreminek/strix-halo. Stack: llama.cpp, IQ4_XS, IQ2_XXS.

tag:computecurve.com,2026:2026-08-31/record/Qwen3.8-Flash-Next/strix-halo/1/58.5 2026-08-31T00:00:00Z

Beats the previous published best of 47.1 for the same model on the same setup, by 24%. Recipe: tomasreminek/strix-halo. Stack: llama.cpp, IQ4_XS, IQ2_XXS.

tag:computecurve.com,2026:2026-08-31/paper/2608.30968 2026-08-31T00:00:00Z

The CogEvol-4B checkpoint ships with code under Apache 2.0 for turning course briefs into educational content. The 4B checkpoint is openly available and can be served locally, though it is a specialized education model. Releases: both.

tag:computecurve.com,2026:2026-08-31/paper/2609.00097 2026-08-31T00:00:00Z

A training-free kernel fuses attention selection and computation, scaling to 256K context length with 11.6x kernel-level speedup. It ships a training-free kernel that can cut long-context decoding latency and memory use on your own GPU. Releases: code.

tag:computecurve.com,2026:2026-08-31/paper/2609.00224 2026-08-31T00:00:00Z

A sub-2-bit PTQ recipe ships with a lookup-table kernel for single-GPU generation. It ships a quantization recipe and a kernel that could cut memory and speed up generation on a single GPU. Releases: code.

tag:computecurve.com,2026:2026-08-31/paper/2609.00355 2026-08-31T00:00:00Z

The one-pass block drafter reads the target VLM's fused vision-language state and ships code for grounded workloads without changing the model. It ships code for lossless speculative decoding on VLMs, letting you run grounded vision-language workloads faster without changing the target model. Releases: code.

tag:computecurve.com,2026:2026-08-30/record/Qwen3.8-Flash-Next/apple-silicon/1/83.06 2026-08-30T00:00:00Z

Beats the previous published best of 12.0 for the same model on the same setup, by 592%. Recipe: Weschera/qwen38-flash-next-omlx-mac. Stack: mlx, oQ.

tag:computecurve.com,2026:2026-08-30/record/Qwen3.8-Flash-Next/rtx-3090/2/135.2 2026-08-30T00:00:00Z

Beats the previous published best of 53.3 for the same model on the same setup, by 154%. Recipe: DominikBucko/qwen38-flash-next-2x3090. Stack: vllm, W4A16, FP8.

tag:computecurve.com,2026:2026-08-30/paper/2608.29934 2026-08-30T00:00:00Z

A parameter-free drop-in adapter trains on ~2.6K MuSiQue 2-hop QA examples and preserves correct answers on evidence-retaining prompts. The released adapter can be dropped into an existing serving stack to cut hallucinations under KV eviction without retraining the base model. Releases: both.

tag:computecurve.com,2026:2026-08-30/paper/2608.30092 2026-08-30T00:00:00Z

Apache-2.0 weights include base and instruction-tuned checkpoints, but the paper releases no code or recipe to reproduce them. You can download and run the weights locally, but the paper ships no code or recipe to reproduce them. Releases: weights.

tag:computecurve.com,2026:2026-08-30/paper/2609.00066 2026-08-30T00:00:00Z

The outlier-companion grouping method ships code and keeps prefill speedup close to RTN. It ships code for a PTQ method that lowers perplexity while keeping decoding memory and prefill speed comparable to RTN. Releases: code.

tag:computecurve.com,2026:2026-08-29/record/GLM-5.3-Flash/dgx-spark/4/102.0 2026-08-29T00:00:00Z

Beats the previous published best of 48.1 for the same model on the same setup, by 112%. Recipe: tonyliu312/GLM-5.3-Flash-DFlash2-TP4-1M-Context. Stack: vllm, NVFP4.

tag:computecurve.com,2026:2026-08-29/record/Qwen3.8-Flash-Next/strix-halo/1/47.1 2026-08-29T00:00:00Z

Beats the previous published best of 33.4 for the same model on the same setup, by 41%. Recipe: drluoto/flash-next-strix-halo. Stack: llama.cpp, UD-IQ4_XS, Q8_0.

tag:computecurve.com,2026:2026-08-29/record/GLM-5.3-Flash/dgx-spark/3/69.0 2026-08-29T00:00:00Z

Beats the previous published best of 64.3 for the same model on the same setup, by 7%. Recipe: FlyCockpit/GLM-5.3-Flash-EXL3-3x-DGX-Sparks. Stack: other, EXL3.

tag:computecurve.com,2026:2026-08-28/record/GLM-5.3/dgx-spark/4/53.32 2026-08-28T00:00:00Z

Beats the previous published best of 26.0 for the same model on the same setup, by 105%. Recipe: tonyd2wild/GLM-5.3-Int4-Int8Mix-TP4-4x-DGX-Spark. Stack: vllm, INT4, INT8.

tag:computecurve.com,2026:2026-08-28/record/GLM-5.3-Flash/dgx-spark/1/43.43 2026-08-28T00:00:00Z

Beats the previous published best of 33.8 for the same model on the same setup, by 28%. Recipe: vcruz305/GLM-5.3-Flash-DFlash2-DGX-Spark-recipe. Stack: llama.cpp, Q2_K, Q4_K_M.

tag:computecurve.com,2026:2026-08-28/record/GLM-5.3-Flash/dgx-spark/2/62.9 2026-08-28T00:00:00Z

Beats the previous published best of 51.0 for the same model on the same setup, by 23%. Recipe: loud1990/GLM-5.3-Flash-NVFP4-MXFP8-2x-DGX-Sparks. Stack: vllm, EXL3, NVFP4.

tag:computecurve.com,2026:2026-08-28/record/Qwen3.8-27B/rtx-3090/1/140.5 2026-08-28T00:00:00Z

Beats the previous published best of 124.8 for the same model on the same setup, by 13%. Recipe: PixelML/Qwen3.8-27B-CMP-170HX. Stack: vllm, W4A16, W8A16.

tag:computecurve.com,2026:2026-08-28/paper/2608.27875 2026-08-28T00:00:00Z

The repo keeps vertical-line tokens and local-window states in high precision while quantizing the rest of attention and KV cache. It ships code for a hybrid-precision attention/KV operator that can cut memory and improve decode efficiency on a single GPU. Releases: code.

tag:computecurve.com,2026:2026-08-28/paper/2608.28458 2026-08-28T00:00:00Z

Released weights for the tuned 2B agent come with a recipe that improves public clemscore from 10.67 to 38.92 while holding aggregate static performance at 44.24. The HF model card links to released weights, so a user can download and run the tuned 2B agent locally. Releases: weights.

tag:computecurve.com,2026:2026-08-27/record/Qwen3.8-27B/rtx-4090/1/339.0 2026-08-27T00:00:00Z

Beats the previous published best of 125.43 for the same model on the same setup, by 170%. Recipe: Anothers/qwen38-27b-4090-48gb-agent-serving. Stack: vllm, W4A16.

tag:computecurve.com,2026:2026-08-27/record/GLM-5.3-Flash/dgx-spark/1/33.8 2026-08-27T00:00:00Z

Beats the previous published best of 18.8 for the same model on the same setup, by 80%. Recipe: Weschera/glm53-flash-one-spark. Stack: llama.cpp, UD-Q2_K_XL.

tag:computecurve.com,2026:2026-08-27/record/DeepSeek-V4-Flash/mi300x/1/280.6 2026-08-27T00:00:00Z

Beats the previous published best of 168.6 for the same model on the same setup, by 66%. Recipe: arjhinety/deepseek-v4-flash-ablit-mi300x. Stack: vllm, FP8, MXFP4.

tag:computecurve.com,2026:2026-08-27/record/Qwen3.8-Flash-Next/dgx-spark/1/50.6 2026-08-27T00:00:00Z

Beats the previous published best of 42.14 for the same model on the same setup, by 20%. Recipe: shantanugoel/qwen38-flash-next-sglang-dgx-spark. Stack: sglang, NVFP4.

tag:computecurve.com,2026:2026-08-26/record/Qwen3-VL-8B-Instruct/dgx-spark/1/27.0 2026-08-26T00:00:00Z

Beats the previous published best of 14.6 for the same model on the same setup, by 85%. Recipe: AI-Architect-Lab-333/dgx-spark-vl-beside-llm. Stack: llama.cpp, Q8_0.

tag:computecurve.com,2026:2026-08-26/record/Qwen3.8-Flash-Next/rtx-pro-6000/1/225.8 2026-08-26T00:00:00Z

Beats the previous published best of 144.563 for the same model on the same setup, by 56%. Recipe: SSHdotCodes/qwen-3.8-flash-next-pro6000. Stack: sglang, NVFP4, BF16.

tag:computecurve.com,2026:2026-08-26/paper/2608.26449 2026-08-26T00:00:00Z

A released harness regenerates every reported number from public data on a laptop, fixing abugida fertility. You can download the released tokenizer and regenerate the reported numbers on a laptop, fixing fertility for abugida languages. Releases: both.

tag:computecurve.com,2026:2026-08-25/paper/2608.24411 2026-08-25T00:00:00Z

Code ships for a proposal-distribution reshaping method that drops into existing serving stacks. It ships code for a speculative decoding method that can be dropped into existing serving stacks to raise token acceptance rates. Releases: code.

tag:computecurve.com,2026:2026-08-24/record/Qwen3.5-27B/apple-silicon/1/52.84 2026-08-24T00:00:00Z

Beats the previous published best of 42.3 for the same model on the same setup, by 25%. Recipe: joonyeonglim/dflash-m4-max-benchmarks. Stack: mlx, INT4.

tag:computecurve.com,2026:2026-08-24/paper/2608.22761 2026-08-24T00:00:00Z

A parameter-free penalty on tokens extending a seen suffix ships as a drop-in across three open-source inference frameworks with no retraining needed. It ships in llama.cpp, ExLlamaV2 and text-generation-webui, so anyone serving open models can enable it to stop loops without retraining. Releases: code.

tag:computecurve.com,2026:2026-08-24/paper/2608.23806 2026-08-24T00:00:00Z

All three checkpoints — the 10B MoE, a dense 3B, and a distilled 480M — are released with vLLM throughput numbers for local serving. All three checkpoints are released and vLLM throughput numbers are directly actionable for local embedding serving. Releases: weights.

tag:computecurve.com,2026:2026-08-24/paper/2608.23816 2026-08-24T00:00:00Z

The recipe ships code and a reproducible setup that avoids search or calibration on a single machine. It ships a reproducible recipe and code that can make quantized LoRA fine-tuning measurably faster on a single machine. Releases: both.

tag:computecurve.com,2026:2026-08-24/paper/2608.23841 2026-08-24T00:00:00Z

A tile layout and CPU streaming engine beat llama.cpp and vLLM on a 32-vCPU server, with weights and a runnable recipe shipped. It ships weights and a CPU inference engine that beats llama.cpp and vLLM on a 32-vCPU server, with a recipe you can run on your own machine. Releases: both.

tag:computecurve.com,2026:2026-08-23/record/Qwen3.8-27B/rtx-pro-6000/1/455.0 2026-08-23T00:00:00Z

Beats the previous published best of 65.54 for the same model on the same setup, by 594%. Recipe: e-accelerate/Qwen3.8-27B-DFlash2-Triton-Blackwell. Stack: sglang, NVFP4, FP8.

tag:computecurve.com,2026:2026-08-22/record/Qwen3.8-27B/rtx-4090/1/125.43 2026-08-22T00:00:00Z

Beats the previous published best of 81.711 for the same model on the same setup, by 54%. Recipe: hugosmoreira/qwen38-27b-rtx4090-windows-benchmark. Stack: llama.cpp, UD-IQ2_XXS, UD-Q2_K_XL.

tag:computecurve.com,2026:2026-08-22/record/Qwen3.8-Flash-Next/dgx-spark/1/42.14 2026-08-22T00:00:00Z

Beats the previous published best of 38.889 for the same model on the same setup, by 8%. Recipe: sixteen-miles-labs/sparklab. Stack: other, NVFP4, DS-FP4.

tag:computecurve.com,2026:2026-08-21/record/Qwen3.8-27B/rtx-3090/1/124.8 2026-08-21T00:00:00Z

Beats the previous published best of 70.0 for the same model on the same setup, by 78%. Recipe: qubex22/qwen38-27b-rtx3090-notes. Stack: vllm, llama.cpp, W4A16.

tag:computecurve.com,2026:2026-08-21/paper/2608.20953 2026-08-21T00:00:00Z

The open-weight 4-bit MXFP4 checkpoint halves the 120B teacher's parameter count and comes with a distillation recipe. It ships a ready-to-run 4-bit checkpoint that halves parameter count and weight memory while matching the bfloat16 source on most benchmarks. Releases: weights.

tag:computecurve.com,2026:2026-08-21/paper/2608.21500 2026-08-21T00:00:00Z

Code and the Qwen3.6-27B-SecOPD checkpoint ship on Hugging Face for on-policy defensive fine-tuning. A released Qwen3.6-27B checkpoint and code let you deploy a defensively fine-tuned model on your own hardware. Releases: both.

tag:computecurve.com,2026:2026-08-20/record/Qwen3.8-27B/rtx-3090/4/242.7 2026-08-20T00:00:00Z

Beats the previous published best of 74.6 for the same model on the same setup, by 225%. Recipe: tonyd2wild/Qwen3.8-27B-DFlash2-4x3090-TP4. Stack: vllm, W4A16.

tag:computecurve.com,2026:2026-08-20/record/Inkling-Small-NVFP4/dgx-spark/2/80.0 2026-08-20T00:00:00Z

Beats the previous published best of 34.0 for the same model on the same setup, by 135%. Recipe: cheller505/gb10-sglang-validation. Stack: sglang, NVFP4.

tag:computecurve.com,2026:2026-08-20/record/Qwen3.8-27B/rtx-5090/1/518.0 2026-08-20T00:00:00Z

Beats the previous published best of 300.0 for the same model on the same setup, by 73%. Recipe: m-sigepon/tidespec. Stack: sglang, NVFP4.