Choose a checkpoint for the job you want to do. The four cells on each card separate the file and compute, local entry point, HTTP service, and client that exercised it. Open the model profile for the command, context evidence and training coverage.
Try two laptop-sized models · Start an agent with Nemotron · Check a live endpoint
Reading the evidence. Local pass means a request completed on this laptop. Recorded pass means CKE retained a result for a specific artifact. A declared profile describes a configuration. Unverified means this guide has no matching acceptance result. CPU entries describe reported hardware. ARM throughput needs a separate measurement for the same weights.
Jump to Nemotron and Deep Agents, Qwen3.8 and Qwen Code, or Gemma4 and Antigravity.
Compare Gemma3 270M IT
- Compute and format Q5_K_M · Commodity x86 nightly lanes.
- Local entry point Text REPL.
- Server Local HTTP pass. One-word reply; 1K build.
- Harness Tool workflow unverified. Run a named client against an identified bundle.
Compare Gemma4 E4B IT
- Compute and format Q4_K_M · Intel AVX2 and AMD Zen 5 development runs.
- Local entry point Text REPL.
- Server Recorded HTTP pass. Chat and tools use distinct variants.
- Harness Antigravity SDK. Four short cases on identified 2K and 32K-capacity tools builds. Test longer inputs and repeated runs.
Compare Qwen2 0.5B Instruct
- Compute and format Q4_K_M · Commodity x86 nightly lanes.
- Local entry point Text REPL.
- Server Local HTTP pass. Qwen2 answered via Responses on the cached 1K bundle.
- Harness Tool workflow unverified. Test a client and a complete tool round trip.
Compare Qwen3 0.6B
- Compute and format Q8_0 · Commodity x86 nightly lanes.
- Local entry point Text REPL.
- Server Local HTTP pass. One-word reply; 1K build.
- Harness Tool workflow unverified. Verify a client on this 0.6B checkpoint.
Compare Qwen3.5 0.8B
- Compute and format Q4_K_M · Commodity x86 nightly lanes.
- Local entry point Text REPL.
- Server Unverified for this artifact. Declared profile qwen-tools-v1.
- Harness No recorded client task. Run one identified request, then a task with a checked result.
Compare Qwen3.5 35B-A3B
- Compute and format Q4_K_M · Intel AVX2/AVX-VNNI and AMD Zen 5 development runs.
- Local entry point Text REPL.
- Server Unverified for this artifact. Declared profile qwen-tools-v1.
- Harness No recorded client task. Run one identified request, then a task with a checked result.
Compare Qwen3.6 27B
- Compute and format Q4_K_M · Intel Ice Lake AVX-512/VNNI x16 measured.
- Local entry point Text REPL.
- Server Unverified for this artifact. Declared profile qwen-tools-v1.
- Harness No recorded client task. Run one identified request, then a task with a checked result.
Compare Qwen3.8 27B dense
- Compute and format Q4_K_M · Ryzen 9 9950X3D Zen 5 artifact runs.
- Local entry point Text REPL.
- Server Recorded HTTP pass. Identified 32K dense bundle.
- Harness Qwen Code. Repeated read, edit and test tasks. Recheck each new quantization or context build.
Compare Nemotron Nano 9B v2
- Compute and format Q4_K_M · x86 runtime and parity lanes.
- Local entry point Text REPL.
- Server Recorded HTTP pass. Explicit tools variant on a 20K-capacity build.
- Harness LangChain Deep Agents. Read, edit, validate and failed-read recovery passed. Test sustained use and cancellation.
Compare GLM4 9B
- Compute and format Q4_K_M · x86 runtime and parity lanes.
- Local entry point Text REPL.
- Server Unverified for this artifact. Check for a service that accepts this artifact.
- Harness No recorded client task. Run one identified request, then a task with a checked result.
Compare Nanbeige 4.1 3B / Llama-family
- Compute and format See checkpoint · Commodity x86 nightly lanes.
- Local entry point Text REPL.
- Server Unverified for this artifact. Check for a service that accepts this artifact.
- Harness No recorded client task. Run one identified request, then a task with a checked result.
Open Nanbeige 4.1 3B / Llama-family
Compare Cohere2 Command R
- Compute and format Q4_K_M · x86 runtime lane.
- Local entry point Text REPL.
- Server Recorded chat pass. Plain chat and streaming on the R7B 2K build.
- Harness LangChain plain chat. The tested template rejects tool requests.
Compare Cohere North Mini Code
- Compute and format Q4_K_M · Ryzen 9 9950X3D AVX-512/VNNI.
- Local entry point Text REPL.
- Server Unverified for this artifact. Check for a service that accepts this artifact.
- Harness No recorded client task. Run one identified request, then a task with a checked result.
Compare Cohere Command A+ 218B/25B-active
- Compute and format W4A4 NVFP4 · Ryzen 9 9950X3D short bring-up.
- Local entry point Specialized text runner.
- Server Unverified for this artifact. Check for a service that accepts this artifact.
- Harness No recorded client task. Run one identified request, then a task with a checked result.
Open Cohere Command A+ 218B/25B-active
Compare Laguna-XS 2.1
- Compute and format Q4_K_M · Ryzen Zen 5 bring-up.
- Local entry point Text REPL.
- Server Unverified for this artifact. Check for a service that accepts this artifact.
- Harness No recorded client task. Run one identified request, then a task with a checked result.
Compare Kimi-VL A3B text decoder
- Compute and format BF16 · Intel AVX2 and Ryzen AVX-512.
- Local entry point Specialized text runner.
- Server Unverified for this artifact. Check for a service that accepts this artifact.
- Harness No recorded client task. Run one identified request, then a task with a checked result.
Compare Muse-Glimmer 30B text decoder
- Compute and format BF16 · Ryzen 9 9950X3D AVX-512 BF16 with oneDNN 3.12 and PyTorch SLEEF.
- Local entry point Specialized text runner.
- Server Recorded tool call. A 32K bundle completed a direct HTTP file-read and continuation.
- Harness Muse Code / Hermes pending. Captured Muse Code input exceeded 32K; tested Hermes required 64K. Retest with sufficient context.
Open Muse-Glimmer 30B text decoder
Compare Instella-MoE 16B-A3B
- Compute and format BF16 · x86 BF16 reference/runtime lane.
- Local entry point Specialized text runner.
- Server Recorded chat pass. Plain chat and streaming on a 4K bundle.
- Harness OpenHands proposed. This variant has no tool protocol. Rebuild and retest prompt admission.
Compare GPT-2
- Compute and format See GPT-2 conversion · Commodity x86 nightly lanes.
- Local entry point Raw continuation.
- Server Unverified for this artifact. Check for a service that accepts this artifact.
- Harness No recorded client task. Run one identified request, then a task with a checked result.
Compare Qwen3-VL 8B
- Compute and format Q4_K_M · Intel and AMD x86 vision/parity lanes.
- Local entry point Image + prompt.
- Server Unverified for this artifact. Check for a service that accepts this artifact.
- Harness No recorded client task. Run one identified request, then a task with a checked result.
Compare Qwen3.6-VL
- Compute and format Q4_K_M · Intel and AMD x86 vision/parity lanes.
- Local entry point Image + prompt.
- Server Unverified for this artifact. Check for a service that accepts this artifact.
- Harness No recorded client task. Run one identified request, then a task with a checked result.
Compare Whisper Tiny, Base, or Small
- Compute and format FP32 · Generated x86 encoder/decoder runtime.
- Local entry point Audio transcription.
- Server Speech HTTP service. Whisper has a dedicated transcription service.
- Harness Audio client. Verify the transcript with a known recording for each size.
Open Whisper Tiny, Base, or Small
Compare Parakeet TDT 0.6B v3
- Compute and format FP32 · Historical Python reference on Ryzen 9 9950X3D: five-minute 7.25% WER, RTF 0.771.
- Local entry point Audio transcription.
- Server Unverified for this artifact. Check for a service that accepts this artifact.
- Harness No recorded client task. Run one identified request, then a task with a checked result.
Compare Cohere Transcribe 2B
- Compute and format Mixed FP16/FP32 · Historical Python reference on Ryzen 9 9950X3D: five-minute RTF 0.220–0.225, about 8.51 GB peak RSS.
- Local entry point Audio transcription.
- Server Unverified for this artifact. Check for a service that accepts this artifact.
- Harness No recorded client task. Run one identified request, then a task with a checked result.
Compare North Micro Vision Instruct
- Compute and format Checkpoint-specific · GPU/ARM coverage unverified.
- Local entry point Image extraction.
- Server Unverified for this artifact. Vision input
- Harness No recorded coding task. Use the vision guide and check extracted fields on a labeled image.
Open North Micro Vision Instruct
Compare Qwen3.8 Flash Next
- Compute and format Q4_K_M · pinned local artifact · Ryzen Zen 5.
- Local entry point Short decode comparison.
- Server Unverified for this artifact. Text decoder
- Harness No recorded coding task. A tool profile is declared. Rebuild the serving bundle and run a real client task.
Compare Kokoro 82M
- Compute and format Pinned voice and weights · Intel i7-14700T.
- Local entry point WAV generation.
- Server Speech HTTP service. Text to speech
- Harness No recorded coding task. A dedicated Kokoro HTTP adapter exists. Check the selected voice, output audio and latency.
Compare Ming3 TTS 0.5B
- Compute and format FP32 compute from BF16 weights · CPU; see model report.
- Local entry point WAV generation.
- Server Unverified for this artifact. Text to speech
- Harness No recorded coding task. Check listening quality and a matching service deployment.
Compare Qwen3 TTS CustomVoice
- Compute and format Pinned CustomVoice session · Intel i7-14700T.
- Local entry point WAV generation.
- Server Unverified for this artifact. Text to speech
- Harness No recorded coding task. Validate the selected language and voice through a service.
Compare Clef Flash decision head
- Compute and format FP32 supplied-feature head · Native CPU.
- Local entry point Local job client.
- Server Unverified for this artifact. Feature scoring
- Harness No recorded coding task. Verify job artifacts and reloaded scores. Text chat is outside this recipe.
Compare RWKV-7
- Compute and format FP32 candidate operations · Whole-model CPU run pending.
- Local entry point Operation tests.
- Server Unverified for this artifact. Candidate
- Harness No recorded coding task. Run a complete checkpoint and verify its outputs.
Compare Gemma4 speculative assistant
- Compute and format Matching draft and backbone · See paired run record.
- Local entry point Pair validation.
- Server Unverified for this artifact. Paired inference
- Harness No recorded coding task. Validate the pair together before measuring speed or serving it.
Open Gemma4 speculative assistant
Compare Inkling 2 Mini
- Compute and format Q8_0 preview · eight-thread CPU trial.
- Local entry point Experimental text generation.
- Server Unverified for this artifact. A preview generated 128 tokens; the expected answer was missing.
- Harness No recorded client task. Verify end-to-end output before advertising a server.
Evidence and coverage
Reviewed model artifacts · Dated harness acceptance · Circuit serving declarations.