Nishi FamilyCompare › LLM Inference and Serving (SOTA)

Nishi vs the Field

LLM Inference and Serving — the full field, measured, sourced.

12 engines, 42 axes, 7 categories. Competitor cells sourced from each project's own docs (researcher-banked 2026-07-09); Nishi cells measured by our gates. Frontier ladder: /compare/llm/frontier

How this is scored. This is a state-of-the-art comparison across the FULL competitor field: quantitative axes carry measured / published numbers (Nishi’s column is our own measurement, competitors are researcher-sourced), grade axes use Best / Yes / Part / No. Every axis carries a source note. No single vanity ‘coverage’ score — the honest picture is per-axis: Nishi leads only on sovereignty + determinism and trails on raw speed, scale, and breadth of features, by design (a research-grade sovereign engine, not a production GPU stack).
CapabilityNishivLLMSGLangTensorRT-LLMllama.cppLMDeployTGIMLC-LLMOllamaExLlamaV2CTranslate2DeepSpeed-MIIAphroditeNotes / source
Speed and scale (quantitative)
Single-stream CPU decode tok/s (0.5B-class)2.40 measuredn/a GPU-firstn/a GPU-firstn/a GPU-only~20-60n/a GPU-firstn/a CPU subpar~10-30~20-60n/a GPU-only~10-40 int8n/a GPU-firstn/a GPU-firstNishi MEASURED 2026-07-09 (Qwen2.5-0.5B, 16T, ~123x cumulative from 51 s/tok); ~20-60 = cited same-class llama.cpp 4-bit CPU, so Nishi is ~8-25x behind same-hardware SOTA
Largest model served (published)0.5B proven670B+ DeepSeek-V3670B+ (96xH100 EP)670B+ Blackwell670B GGUF1TB Intern-S1-Pro180B Falcon70B-class670B library120B-class multi-GPU20B-class enc-dec530B MT-NLG670B+ vLLM-parityfrom each project README/model table (LMDeploy lists Intern-S1-Pro 1TB); Nishi = Qwen2.5-0.5B end-to-end proven
Weight-quant formats4 GGUF reads + Q8_0 repack10+ (FP8 MXFP4 NVFP4 INT8/4 GPTQ AWQ GGUF...)5+ (FP4 FP8 INT4 AWQ GPTQ)6+ (FP8 FP4 INT4 AWQ)25+ (1.5- to 8-bit GGUF)6+ (AWQ GPTQ W4A16 MXFP4 KV int8/4)7 (bnb GPTQ EETQ AWQ Marlin fp8 NF4)3-4 group-quantGGUF setEXL2 2-8 bpw3 (INT8 INT16 AWQ)2-310+ vLLM-derivedcounts from each README quant list; Nishi = Q4_K/Q5_0/Q6_K/Q8_0 read + in-memory Q8_0 repack (measured arc)
Model architectures2 (Qwen2-class GGUF + nofloat custom)200+wide (LLM embed reward diffusion)50+150+ GGUF60+ LLM + 40+ VLM20+ optimized40+100s libraryLlama-family+20+ (enc-dec incl Whisper)20+vLLM-parityvLLM README claims 200+; LMDeploy README model table; Nishi loads Qwen2-class GGUF plus its own trained nofloat transformer
Hardware targets1 (x86-64 ELF)10+ (NV AMD x86 ARM PPC TPU Gaudi Ascend Apple...)6+ (NV AMD Xeon TPU Ascend)1 vendor (NVIDIA all gens)10+ (CPU CUDA Metal Vulkan HIP SYCL MUSA CANN OpenCL RPC)3 (NV V100+ Ascend Jetson)6 (NV AMD Inferentia Intel Gaudi TPU)8+ (CUDA ROCm Vulkan Metal WebGPU iOS Android)4 (Mac Win Linux; NV AMD)2 (NVIDIA ROCm)2 (x86 CPU NVIDIA)1-2 (NVIDIA)4+ (NV AMD CPU Win)backend lists from each README; llama.cpp is the portability king of the natives, MLC of browser/mobile
Production scale (self-reported)1 home box (RTX5080 NAS, live site)2000+ contributors, industry standard400k+ GPUs, trillions tok/dayNVIDIA enterprise stackubiquitous local standardInternLM productionHF prod, now maintenance modebrowser+mobile nichedominant local runnerenthusiast standardprod MT/Whisper stacksdormant since ~2024community serving (Kobold lineage)self-reported README claims (SGLang: 400k GPUs; TGI: CAUTION maintenance mode)
KV cache and memory
Paged / blocked KV cachePartBestYesYesPartYesYesYesPartYesNoYesYesvLLM invented PagedAttention (SOSP 2023); Nishi nx_kvcache BUILT + gate GREEN (paged==contiguous bit-exact, fork+COW) but not yet wired into serving; ExLlamaV2 paged via flash-attn; llama.cpp unified cache
Continuous batchingNoBestBestYesYesYesYesYesPartPartPartYesYesSGLang zero-overhead scheduler + vLLM set the bar; LMDeploy persistent batch; MII Dynamic SplitFuse; CT2 dynamic offline batching; Nishi serves one request at a time (top gap)
Prefix caching / reuseNoYesBestYesPartYesPartPartPartYesNoNoYesSGLang RadixAttention (5x claim, banked blog list); TRT-LLM KV Cache Reuse; LMDeploy automatic prefix caching; llama.cpp prompt cache
Chunked prefillYesYesYesYesYesYesYesPartPartYesNoYesYesNishi WIRED + measured 2026-07-09 (one m=n forward, ~17x fewer prefill passes, TTFT collapse); LMDeploy dynamic split-and-fuse; MII SplitFuse
KV-cache quantizationNoYesYesYesYesYesNoPartYesBestNoNoYesExLlamaV2 Q4 cache pioneer (README); LMDeploy online int8/int4 KV (README); llama.cpp q8/q4 KV flags
Speed techniques
Fused quantized-read GEMMPartBestYesBestYesYesYesYesYesYesYesYesYesMarlin-class W4A16 is the SOTA bar; Nishi CPU-Marlin: __f32_i8dot32 SSE intrinsic blessed, Q8_0 path 10.2x measured, Q5/Q4 unpack proven compute-bound on CPU (honest partial)
Speculative decodingPartYesBestYesYesPartYesNoNoYesNoNoYesSGLang DFlash/Spec-V2 (2026-06 blog = current frontier); TGI claims ~2x; Nishi BUILT prompt-lookup + bit-identical gate, 58 pct accept, wall 0.94x = no net win yet (honest)
Compiled / graph executionPartYesYesBestYesYesYesBestYesPartYesYesYesTRT engine compilation and MLC whole-model TVM compilation are the two compilation identities; vLLM CUDA graphs + torch.compile; Nishi is AOT-native by construction but has no GPU graphs
Optimized attention kernelsPartBestYesYesYesYesYesYesYesYesPartYesYesvLLM README lists FlashAttention FlashInfer TRTLLM-GEN FlashMLA Triton; LMDeploy FlashMLA; Nishi has own GQA offset-causal attention, scalar+SSE, no flash-class tiling
Parallelism and distribution
Tensor parallelismNoYesYesYesPartYesYesPartNoYesYesYesYesTGI NCCL TP (README); CTranslate2 tensor parallelism (README); llama.cpp row/layer split
Pipeline parallelismNoYesYesYesYesYesPartNoNoNoNoPartYesvLLM/SGLang/TRT READMEs; llama.cpp RPC layer split
Expert parallelism / MoENoYesBestYesYesYesPartPartYesPartNoPartYesSGLang large-scale EP (96xH100 + GB200 blogs) leads; LMDeploy DeepEP/eplb (README); Nishi has no MoE (remaining census opportunity)
Prefill/decode disaggregationNoYesBestYesNoYesNoNoNoNoNoNoPartSGLang GB200 P/D blogs (3.8x prefill 4.8x decode); LMDeploy via DLSlime+Mooncake (README); TRT via Dynamo
Multi-node servingNoYesYesYesYesYesYesNoNoNoPartYesYesllama.cpp RPC; LMDeploy proxy server multi-machine (README); MII replicas + load balancing (README)
CPU multithread scalingYesPartPartNoBestNoNoPartYesNoYesNoNoNishi MEASURED 7.34x q4k matmul on 8T + 16T decode (sovereign pool, no pthread); llama.cpp AVX/AMX/NUMA is the CPU king; CT2 MKL/oneDNN/Ruy
Model I/O and sampling
GGUF loadingYesYesPartNoBestNoNoNoYesNoNoNoYesllama.cpp owns the format; Nishi nx_gguf loader + tensor-type census (Q4_K/Q5_0/Q6_K/Q8_0/F32) measured; Ollama/Aphrodite GGUF native
Own tokenizer (no external lib)YesNoNoNoYesNoNoNoYesNoNoNoNoNishi sovereign BPE verified == Qwen2 digit-split behavior (nx_tok_probe); llama.cpp own tokenizer; the rest ship HF tokenizers
Sampling suitePartYesYesYesBestYesYesYesYesYesYesYesYesllama.cpp deepest (min-p mirostat DRY XTC ecosystem); TGI logits warpers (README); Nishi greedy + temperature + top-k measured, no top-p/penalties yet
Structured output (JSON/grammar)NoYesBestYesYesYesYesYesYesPartNoNoYesSGLang compressed-FSM 3x JSON decoding (banked blog list); llama.cpp GBNF grammars; vLLM xgrammar/guidance (README); Ollama JSON schema
Tool / function callingNoYesYesYesYesYesYesYesYesPartNoNoYesvLLM tool calling + reasoning parsers (README); LMDeploy TOOLS CALLING (README); Ollama tool calling (README)
Multimodal (vision) servingNoYesYesYesYesBestPartPartYesPartNoNoYesLMDeploy VLM table is the largest (InternVL home, README); llama.cpp multimodal; Ollama vision models
Embeddings / rerank servingPartYesYesPartYesPartPartNoYesNoPartNoYesllama.cpp embedding + rerank (README); SGLang e5/gte/reward (README); Nishi sovereign SGNS embeddings power live in-corpus search but expose no embed-API endpoint
In-stack training / autogradYesNoPartNoPartNoNoNoNoNoNoNoNoNishi sovereign autograd + stable Q16 integer training recipe (trained real models: reader, neural image codec); SGLang = RL rollout backend for external trainers; the field is inference-only
Serving and API
OpenAI-compatible APINoYesYesYesYesYesYesYesYesYesNoPartYesuniversal in the field (each README); Nishi serve organ is bespoke HTTP + MCP-first plane instead - an honest drop-in-adoption gap
Anthropic Messages APINoYesPartNoNoNoNoNoNoNoNoNoNovLLM README explicitly ships Anthropic Messages API alongside OpenAI + gRPC
gRPC APINoYesNoYesNoNoYesNoNoNoNoYesNoTGI Rust gRPC core (README); MII gRPC (README); TRT via Triton
Agent-native MCP tool surfaceYesNoNoNoNoNoNoNoNoNoNoNoNoNishi live sovereign MCP plane (compare/status/search tools served to agents today); absent from every banked engine README
Token streamingNoYesYesYesYesYesYesYesYesYesPartYesYesTGI SSE (README); Ollama stream (README); Nishi serve returns full completions (gap)
Metrics / tracingNoYesYesYesYesPartBestNoPartNoNoPartYesTGI Prometheus + OpenTelemetry distributed tracing (README) = the production bar; Nishi has logs only
Turnkey local installPartPartPartNoYesPartPartYesBestPartYesPartPartOllama one-command install is the consumer bar (README); Nishi = sovereign ELF drop, zero deps, but self-hosted only
Sovereignty and determinism (Nishi lane)
Zero third-party runtime depsBestNoNoNoYesNoNoPartPartNoPartNoNoNishi: own language + compiler to x86 ELF, own HTTP/TLS, no libc/BLAS/PyTorch/CUDA; llama.cpp minimal C/C++ but on OS toolchain; the rest sit on PyTorch/CUDA/HF
Bit-exact deterministic decodeBestNoNoNoPartNoNoNoPartNoPartNoNoNishi gates prove bit-identical decode (spec==plain, paged==contiguous, cross-run); batched GPU float engines are nondeterministic by construction; llama.cpp reproducible only single-config
Model trained WITHOUT floatBestNoNoNoNoNoNoNoNoNoNoNoNoNishi integer-only Q16 training (grad-clip + LR-decay recipe, measured) - no engine in the field claims this
Full-stack auditabilityBestNoNoNoYesNoNoPartPartPartPartNoNoNishi: every line from compiler to syscall is in-house readable; llama.cpp compact C++ auditable but sits on drivers/CUDA beneath
Offline / air-gappedBestPartPartPartYesPartPartYesYesYesYesPartPartNishi never phones home by construction; llama.cpp/Ollama fully local (model pull optional); HF-hub-default engines can be configured offline
Actively developed (2026)YesBestBestYesBestYesNoPartYesPartPartNoYesfrom banked READMEs: TGI = CAUTION maintenance mode recommending vLLM/SGLang; MII dormant; vLLM/SGLang/llama.cpp are the active frontier
12 competitors42 axes6 quantitativeNishi Best on 5
Honest verdict. Measured across the full field, Nishi is a research-grade sovereign inference engine, not a production stack: measured 2.40 tok/s CPU decode (0.5B) vs a cited ~20-60 tok/s for llama.cpp-class 4-bit on the same hardware class (~8-25x behind same-hardware SOTA) and orders of magnitude behind batched GPU aggregate; it serves a 0.5B model where the field serves 200B-1T, with no GPU, no batching, no multi-node. Speculative decoding and paged-KV are built + bit-exact gated but not yet serving wins; chunked prefill IS wired. Its genuine, unique wins are structural — the whole stack is ours bits-up (own compiler, zero PyTorch / CUDA / BLAS), decode is bit-exact deterministic, and it has trained a model without float — claims no engine in the field makes. Next rungs: batched serving, wire paged-KV + speculation, GPU.

Generated by nx_swcompare_sota from knowledge/compare/llm.sota — quantitative axes measured/sourced; researcher-fed (nx_swcompare_research). Zero JS, zero trackers.