01 · THE MODEL BENCH

The Model Bench

Eleven open models, every one loaded and ready on local hardware. Zero cloud API calls, zero waitlists.

Model table last updated September 2026.

Qwen3.8-27B NVFP4 · benchmark reference ≈86 tok/s

DEFAULT TRIALONE SESSION · TWO ACCESS PATHS

Qwen3.8-27B BF16 / FP8 / NVFP4 W4A4

Every trial session connects to a live local machine running Qwen3.8-27B, accessible through two distinct paths:

Access Path One · Browser

Open WebUI

Open WebUI in the browser. No setup. Good for judging output quality, prompt behavior, and document retrieval (RAG) on your own files.

Access Path Two · API

OpenAI-Compatible Endpoint

An OpenAI-compatible endpoint on the same machine. A base URL, a scoped key, and the served model name. Point your existing tooling at it and change nothing else.

What the endpoint is for

A chat window tells you whether you like the output. It cannot tell you whether the model holds up inside your systems. The endpoint can.

Point your evaluation harness at it and run your own test set. Wire it into a coding tool by changing one base URL. Exercise tool calling and structured output the way your agents actually use them. Generate concurrent load and watch what happens. Run the same prompts you already run against a hosted API and compare directly.

It is OpenAI-compatible, so most tooling needs one configuration change and no code.

Trial machines may be shared between sessions. The API session is for output quality and integration testing. Representative throughput, latency, and concurrency numbers come from the paid Private AI Assessment on a dedicated machine.

Request Trial Access

Session Parameters

What You Receive on Session Start

On session start you receive:

A base URL
A scoped API key, valid for the session window only
The served model name string
The available endpoints: /v1/chat/completions, /v1/completions, /v1/models
The key is revoked when the session ends.

Example Request

curl https://<your-session-host>.gpunexus.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <scoped-api-key>" \
  -d '{
    "model": "qwen3.8-27b",
    "messages": [
      {"role": "user", "content": "Analyze this codebase structure."}
    ]
  }'

Model	Class	Context	Min. Hardware	Engines	Precision & Footprint	Use Case & Capabilities
GLM-5.3-FlashFRONTIER	MoE 320B / 18B active	1M	Dual Spark	vLLM, SGLang	FP8 ~331GB NVFP4 ~170GB	Z.ai open frontier model. Agentic coding and reasoning at a fraction of GLM 5.2's footprint. 4.4x smaller KV cache makes 1M-context serving realistic on-prem.
Qwen3.8-Flash-NextFRONTIER	MoE ~180B / 6B active	262K→1M	Single Spark	vLLM, SGLang	4-bit ~93–111GB FP8 ~173GB	Beats Claude Opus 4.6 on SWE-bench Pro (62.5 vs 53.4). 6B active parameters plus host-memory n-gram offload brings 180B-class quality to Spark-tier memory. Commercial MaaS threshold applies. We confirm license fit during assessment.
GLM 5.2FRONTIER	MoE 743B / 40B active	1M	Quad Spark	vLLM, SGLang	NVFP4~375GB	Zhipu frontier open model. Strong agentic coding and tool use with 1M context under MIT license.
DeepSeek-V4-Flash-0731FRONTIER	MoE 284B / 13B active	1M	Dual Spark	vLLM, SGLang	NVFP4~142GB	Frontier-grade open reasoning, 1M context, MIT license. Runs on dual Spark link or 2x RTX PRO.
Qwen3.8-27BDEFAULT TRIAL	Dense 27.78B	262K→1M	Single Spark	SGLang, vLLM	BF16 ~56GB FP8 ~28GB / NVFP4	Dense 27.78B hybrid decoder with 3:1 linear-to-full attention. Natively multimodal (text, image, video) with multi-token prediction head. Official BF16 and FP8 checkpoints.
Qwen3.6 35B A3B	MoE 35B / 3B active	128K	Single Spark	vLLM, SGLang	BF16~70GB	Efficient active-parameter path. More model than the 27B at lower cost per token.
Qwen3.5 122B A10B	MoE 122B / 10B active	128K	Single Spark	vLLM, SGLang	NVFP4~61GB	Large-model quality on a single box.
Gemma 4 31B	Dense 31B	256K	Single Spark	vLLM, SGLang	BF16~62GB	Strong multilingual and long-document work. 256K context.
Mistral Medium 3.5	Dense 128B	256K	Single Spark	vLLM, SGLang	NVFP4/Q4~64GB	Frontier-class coding and agentic comparison point. 256K context.
Nemotron 3 Super	MoE 120B / 12B active	128K	Single Spark	vLLM, SGLang	NVFP4~60GB	NVIDIA's reasoning line, pretrained natively in NVFP4.
GPT-OSS 120B	MoE 117B / 5B active	128K	Single Spark	vLLM, SGLang	MXFP4~63GB	Open-weight 120B that fits a single box.

GLM-5.3-Flash

FRONTIER

MoE 320B / 18B active · FP8 ~331GB / NVFP4 ~170GB

Context: 1MMin: Dual Spark

Z.ai's newest open frontier model. Frontier-adjacent agentic coding and reasoning at a fraction of GLM 5.2's footprint. The 4.4x smaller KV cache makes 1M-context serving realistic on-prem.

Engines: vLLM, SGLangLicense: MIT

Qwen3.8-Flash-Next

FRONTIER

MoE ~180B / 6B active · 4-bit ~93–111GB / FP8 ~173GB

Context: 262K→1MMin: Single Spark

Beats Claude Opus 4.6 on SWE-bench Pro (62.5 vs 53.4 per Alibaba). The 6B active parameter design plus host-memory n-gram offload makes 180B-class quality realistic on Spark-tier memory.

Commercial MaaS threshold applies. We confirm license fit during assessment.

Engines: vLLM, SGLangLicense: Qwen Community 1.0

GLM 5.2

FRONTIER

MoE 743B / 40B active · NVFP4 ~375GB

Context: 1MMin: Quad Spark

Zhipu's frontier open model, on-premise. Strong agentic coding and tool use with 1M context under MIT license.

Engines: vLLM, SGLangLicense: MIT

DeepSeek-V4-Flash-0731

FRONTIER

MoE 284B / 13B active · NVFP4 ~142GB

Context: 1MMin: Dual Spark

Frontier-grade open reasoning, 1M context, MIT license. Runs on dual Spark link or 2x RTX PRO 6000.

Engines: vLLM, SGLangLicense: MIT

Qwen3.8-27B

DEFAULT TRIAL

Dense 27.78B · BF16 ~56GB / FP8 ~28GB / NVFP4

Context: 262K→1MMin: Single Spark

Dense 27.78B parameter hybrid decoder with Gated DeltaNet linear attention on 48 of 64 layers and full attention on 16 layers (3:1 ratio). Natively multimodal (text, image, video). Built-in multi-token prediction (MTP) head. Official BF16 and FP8 checkpoints.

Engines: SGLang, vLLMLicense: Apache 2.0

Qwen3.6 35B A3B

MoE 35B / 3B active · BF16 ~70GB

Context: 128KMin: Single Spark

Efficient active-parameter path. More model than the 27B at lower cost per token.

Engines: vLLM, SGLangLicense: Apache 2.0

Qwen3.5 122B A10B

MoE 122B / 10B active · NVFP4 ~61GB

Context: 128KMin: Single Spark

Large-model quality on a single box.

Engines: vLLM, SGLangLicense: Apache 2.0

Gemma 4 31B

Dense 31B · BF16 ~62GB

Context: 256KMin: Single Spark

Strong multilingual and long-document work. 256K context.

Engines: vLLM, SGLangLicense: Gemma Terms

Mistral Medium 3.5

Dense 128B · NVFP4/Q4 ~64GB

Context: 256KMin: Single Spark

Frontier-class coding and agentic comparison point. 256K context.

Engines: vLLM, SGLangLicense: Apache 2.0

Nemotron 3 Super

MoE 120B / 12B active · NVFP4 ~60GB

Context: 128KMin: Single Spark

NVIDIA's reasoning line, pretrained natively in NVFP4.

Engines: vLLM, SGLangLicense: NVIDIA Open License

GPT-OSS 120B

MoE 117B / 5B active · MXFP4 ~63GB

Context: 128KMin: Single Spark

Open-weight 120B that fits a single box.

Engines: vLLM, SGLangLicense: MIT

Some frontier open models exceed single-site deployment, for example Kimi K3 at 2.8T total parameters and roughly 1.56TB of weights. We size multi-node builds for these on request.

Test Any Model on Your Prompts

We configure a private session with your preferred model and context length so you can benchmark latency, first-token speed, and output quality.

Get 30-Minute Access Request Full Workload Assessment