GPT-OSS 120B vs GLM-5.1 vs DeepSeek V4: Which Open-Weight Model Wins in 2026?

TL;DR

• For coding on a single GPU: GPT-OSS 120B (fits on 1x H100 SXM5 with MXFP4, from $2.20/hr on Spheron as of 16 Oct 2026)
• For highest coding quality: GLM-5.1 (Chatbot Arena Elo 1467, needs 4-8x H200 SXM5)
• For 1M context and agentic reasoning: DeepSeek V4-Pro (1.6T MoE, needs 8x H200) or the lighter V4-Flash (284B MoE, 2x H200 or 4x H100)

Costs, VRAM requirements, and full benchmarks below.

Three open-weight frontier models are deployable as of April 2026: GPT-OSS 120B, GLM-5.1, and DeepSeek V4. The hardware cost spread between them is enormous. GPT-OSS 120B fits on a single H100 with MXFP4 quantization. GLM-5.1 needs at minimum 4x H200 for INT4 or 8x H200 for FP8. DeepSeek V4-Pro needs a full 8x H200 node, while V4-Flash fits 2x H200 or 4x H100. Picking the wrong model for your budget means either overpaying for capacity you don't need, or under-serving your workload with hardware that can't hold the weights.

This post puts all three models side by side: architecture, VRAM requirements, inference throughput, quality benchmarks, and cost per million tokens at current Spheron pricing. Moonshot's 2.8T Kimi K3 has since raised the ceiling well above all three; its VRAM and cluster sizing requirements are a different class of problem.

TL;DR

Model Architecture Comparison

† GPT-OSS 120B MXFP4 weight size: the ~60GB estimate covers expert weights compressed to 4-bit precision. MXFP4 is an Open Compute Project (OCP) industry standard; native hardware acceleration for it requires Blackwell GPUs (B200, GB200). On H100, the runtime falls back to software emulation of MXFP4 or MXFP8, which reduces throughput compared to native Blackwell. See FP4 quantization on Blackwell for details.

¶ GLM-5.1 INT4 estimate of ~377GB is theoretical minimum at 4-bits per parameter (754B × 0.5 bytes). AWQ/GPTQ INT4 in practice may be slightly larger due to block metadata. FP8 at ~754GB assumes 1 byte per parameter, excluding KV cache and activation overhead. At 32K context length with moderate batch sizes, expect 10-15% additional VRAM for KV cache.

‡ DeepSeek V4 weight sizes are the native release checkpoints, which store the MoE experts in FP4 and everything else in FP8: ~865GB for V4-Pro, ~160GB for V4-Flash. At runtime, add 15-20% for activations and framework buffers, plus KV cache. Upcasting to pure FP8 roughly doubles the expert weights (V4-Pro grows to ~1.6TB), and INT4 re-quants don't shrink either model below its native checkpoint. Both models are MIT-licensed. See the GPU memory requirements guide and the GPU requirements cheat sheet 2026 for VRAM planning details.

All three models are MoE architectures. The active parameter count (the column that drives per-token compute) ranges from ~5.1B (GPT-OSS 120B) to 49B (DeepSeek V4-Pro). Total parameter count sets the memory floor. For a deep dive on why you pay for storage even when only a fraction of parameters are active per token, see the MoE inference optimization guide.

GPU Hardware Requirements and Quantization

GPT-OSS 120B

MXFP4 is a 4-bit floating-point format defined by the Open Compute Project (OCP) industry standard. NVIDIA Blackwell GPUs (B200, GB200) support it natively; NVIDIA also has a proprietary variant called NVFP4. On H100, vLLM falls back to software emulation at reduced throughput; you still get the memory savings but not the hardware speedup. For production on H100, FP8 at 2x H100 may give better throughput per dollar than MXFP4 on 1x H100, depending on your batch size. Link to the full setup guide: deploy GPT-OSS on GPU cloud.

GLM-5.1

GLM-5.1 FP8 is the practical production configuration. At 32K context length with 16 concurrent requests, expect roughly 100-140GB additional VRAM for KV cache on top of the ~754GB weight footprint, so 8x H200 (1,128GB total) is tight at high concurrency. AWQ INT4 on 5x A100 80GB is viable for lower-concurrency deployments where quality loss from aggressive quantization is acceptable. Full setup: deploy GLM-5.1 on GPU cloud.

Z.ai's June 2026 update, GLM-5.2 on Spheron, reduces total parameter count to 744B and expands context to 1M tokens with a coding-first focus.

DeepSeek V4

V4-Pro does not fit on 8x H100 (640GB) or 4x H200 (564GB), and re-quantizing to INT4 doesn't help, since INT4 files come out around 0.6 bytes per parameter, no smaller than the native checkpoint. DeepSeek V4 with expert parallelism requires that the GPU cluster has high-bandwidth interconnect. On 8x H200 SXM5, NVLink handles the expert routing traffic efficiently. A cluster without NVLink (e.g., H100 PCIe) will bottleneck on inter-GPU communication during MoE routing and may not reach useful throughput even with enough aggregate VRAM. Full setup: deploy DeepSeek V4 on GPU cloud. See also the MoE inference optimization guide for expert parallelism configuration details.

Inference Benchmarks

Throughput (Tokens/Second)

Benchmark figures below are estimates based on published vLLM benchmarks and model architecture characteristics. Batch size: 50 concurrent requests, input/output sequence of 512/256 tokens.

These throughput estimates could not be confirmed from official model cards or technical reports for the April 2026 versions. Treat them as directional and run your own vLLM benchmarking on your target hardware before capacity planning. Benchmark methodology and inference engine choice affect numbers significantly; see the vLLM vs TensorRT-LLM vs SGLang comparison for the impact on throughput.

Per-GPU throughput is where GPT-OSS 120B wins decisively. At ~600-800 tokens/sec on a single H100, it outperforms GLM-5.1 on a normalized per-GPU basis by a factor of 5-8x, because GLM-5.1 distributes computation across 8 GPUs to produce similar or lower aggregate throughput. The DeepSeek V4 rows are left unestimated because the earlier figures were sized for pre-release specs that turned out to be wrong; measure V4-Pro and V4-Flash on your own hardware.

TTFT (Time to First Token)

TTFT figures are estimates; actual values depend heavily on prompt length, KV cache state, and batching configuration. For production latency tuning, see the vLLM vs TensorRT-LLM vs SGLang benchmarks post.

Throughput per Dollar

This table is a worked example frozen at the Spheron on-demand rates recorded on 19 Apr 2026 ($4.41/hr for the H100 SXM5 instance, $4.54/hr per H200 SXM5). Today Spheron lists H100 at $2.64/hr and H200 at $4.80/hr per GPU on-demand, as of 16 Oct 2026. Formula: ($/hr / 3600) / (tokens/sec / 1,000,000) = $/M tokens. Substitute current pricing and your measured throughput to recalculate.

The Est. throughput column uses unverified estimates from the benchmarks section above. The resulting $/M tokens figures are also estimates; run your own throughput measurement before using these for budget planning.

Pricing fluctuates based on GPU availability. Spheron rates above are live as of 16 Oct 2026; the cost-per-token table is a worked example at the 19 Apr 2026 rates stated above, and other providers reflect their most recent published rates and may have changed. Check current GPU pricing → for live rates.

Quality Benchmarks

Coding (SWE-bench, HumanEval)

¹ GPT-OSS 120B SWE-bench: figure could not be confirmed from the official OpenAI model card (OpenAI does not publish SWE-bench Pro directly). The 94.6% relative-to-Claude figure is from the GLM-5.1 deployment guide sourced from Z.ai release notes. GPT-OSS figure is an estimate; verify against your own eval before using for production selection.

² GPT-OSS 120B HumanEval: OpenAI release notes for GPT-OSS (August 2025) reference strong coding performance but do not cite a specific HumanEval number in the published materials. Treat as unverified.

³ GLM-5.1 HumanEval percentage: the "94.6% of Claude Opus 4.6" metric is from Z.ai's SWE-bench Pro evaluation, not standard HumanEval. These are different benchmarks. The column header reflects the common metric used; source data is Z.ai's April 2026 release notes.

⁴ DeepSeek V4 benchmark scores: this comparison was first written before V4's April 2026 release, when only pre-release claims existed, so no V4 scores are listed here. Check DeepSeek's official V4-Pro and V4-Flash model cards and run your own eval; the DeepSeek V4 deployment guide covers both models.

General Reasoning (MMLU-Pro)

⁵ GPT-OSS 120B MMLU-Pro: OpenAI's release materials reference competitive MMLU performance but do not publish MMLU-Pro scores directly. Treat as unverified.

⁶ GLM-5.1 Chatbot Arena Elo 1467 is from the official Chatbot Arena leaderboard as of April 2026, sourced from the GLM-5.1 deployment guide. The Elo reflects API-served GLM-5.1; self-hosted FP8 quality may differ slightly from the API endpoint quality.

For all benchmark figures: these numbers are starting points. Run your own evaluation on your specific task distribution before making hardware procurement decisions.

Cost-Per-Token Analysis

On-Demand Pricing

This is a worked example frozen at the rates recorded on 19 Apr 2026: $4.41/hr for the H100 SXM5 instance and $4.54/hr per H200 SXM5, yielding $36.33 for 8x, $18.16 for 4x and $9.08 for 2x, sourced from the H200 rental page. As of 16 Oct 2026, 8x H200 runs $38.40/hr on-demand. Confirm current rates at the pricing page before finalizing budget estimates.

The Est. $/M tokens figures use the unverified throughput estimates from the benchmarks section above. Substitute your measured throughput into the formula ($/hr / 3600) / (tokens/sec / 1,000,000) = $/M tokens for accurate cost modeling.

Pricing fluctuates based on GPU availability. Spheron rates above are live as of 16 Oct 2026; the cost-per-token table is a worked example at the 19 Apr 2026 rates stated above, and other providers reflect their most recent published rates and may have changed. Check current GPU pricing → for live rates.

Spot vs On-Demand

Spheron H200 spot runs $3.32/hr per GPU ($26.56/hr for an 8-GPU node) against $4.80/hr on-demand, as of 16 Oct 2026. Spot fits batch inference and offline pipelines on GLM-5.1's multi-GPU H200 configurations, and the GLM-5.1 deployment guide documents spot details for H200 configurations. For single-GPU GPT-OSS 120B, H100 spot is listed at $2.20/hr. Spot capacity can be reclaimed without notice.

DeepSeek V4-Pro runs on 8x H200, so the same H200 spot options apply to it as to GLM-5.1. Spot inventory changes with partner supply, so check the pricing page for up-to-date spot availability on H100 and H200 before planning a spot-based pipeline for DeepSeek V4 on either 4x H100 or 2x H200.

For workloads where uptime matters, on-demand is the safer choice. For nightly batch jobs, offline evaluation runs, or asynchronous inference queues, compare the 8x H200 spot and on-demand node rates above on the day you launch.

API Pricing Comparison

Self-hosting becomes cost-effective once your token volume is high enough to justify leaving the GPU on. For GPT-OSS 120B at $2.64/hr on-demand, the crossover against a $1-5/M token API depends on utilization. Running the GPU only when needed (spinning up per job), self-hosting beats most API tiers starting around 10M tokens/month. At continuous 24/7 operation, the break-even rises to hundreds of millions of tokens per month.

For GLM-5.1 and DeepSeek V4 on 2x to 8x H200 ($9.60/hr to $38.40/hr on-demand), the same logic applies but the GPU cost is higher, so higher token volumes are required to break even. See the on-premise vs GPU cloud cost breakdown and the GPU cloud pricing comparison 2026 for the full break-even analysis.

Deployment Guide

Deploy GPT-OSS 120B on Spheron

# Single H100 SXM5 with MXFP4
docker run --gpus all --ipc=host -p 8000:8000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  vllm/vllm-openai:latest \
  --model openai/gpt-oss-120b \
  --quantization mxfp4 \
  --gpu-memory-utilization 0.92 \
  --max-model-len 32768

Note: MXFP4 hardware acceleration only applies on Blackwell GPUs (B200, GB200). On H100, vLLM uses software emulation for MXFP4 weight loading, which reduces throughput. If you need maximum throughput on H100, use FP8 with 2x H100 instead. See FP4 quantization on Blackwell GPUs for hardware requirements. Full deployment guide: deploy GPT-OSS on GPU cloud. If your deployment leans on native MXFP4 at rack scale, Spheron has GB200 rack-scale capacity you can reserve today: enter your GPU count, timeline, and workload on the form and the team confirms availability and reaches out within a business day.

Deploy GLM-5.1 on Spheron (8x H200, FP8)

python -m vllm.entrypoints.openai.api_server \
  --model ./glm5-1-fp8 \
  --served-model-name glm5-1-fp8 \
  --tensor-parallel-size 8 \
  --quantization fp8 \
  --enable-expert-parallel \
  --max-model-len 32768 \
  --port 8000

Download weights first: huggingface-cli download zai-org/GLM-5.1-FP8 --local-dir ./glm5-1-fp8. The zai-org/GLM-5.1-FP8 model ID is from Z.ai's official Hugging Face org as of April 2026; verify the ID hasn't changed at the Hugging Face repo before running. Full deployment guide: deploy GLM-5.1 on GPU cloud.

Deploy DeepSeek V4-Pro on Spheron (8x H200)

python -m vllm.entrypoints.openai.api_server \
  --model deepseek-ai/DeepSeek-V4-Pro \
  --tensor-parallel-size 8 \
  --enable-expert-parallel \
  --max-model-len 65536 \
  --port 8000

The V4-Pro checkpoint carries its own FP4/FP8 quantization config, so no --quantization flag is needed, and its ~865GB of weights will not load on 8x H100. For V4-Flash, swap in deepseek-ai/DeepSeek-V4-Flash with --tensor-parallel-size 2 on 2x H200 or --tensor-parallel-size 4 on 4x H100. Full deployment guide: deploy DeepSeek V4 on GPU cloud.

For users running vLLM Model Runner V2 (MRV2), see the vLLM MRV2 deployment guide for configuration differences. Model quick-start configs for all three models are also at docs.spheron.ai/quick-guides/llms/.

Decision Matrix

If you're unsure where to start, deploy GPT-OSS 120B first. It runs on a single H100, costs $2.64/hr on-demand, and gives you a working OpenAI-compatible endpoint without the multi-GPU orchestration complexity of GLM-5.1 or DeepSeek V4. Once you have a baseline and know your token volume and quality requirements, you can decide whether GLM-5.1's coding quality or DeepSeek V4's 1M context window justifies the 2-8x increase in GPU cost. For teams also evaluating Moonshot AI's models, you can deploy Kimi K2.6 for agentic-swarm workloads (300-agent coordination) on a similar 8×H200 hardware setup. Alibaba's Qwen3.7 Max launched in May 2026 but is closed-weight and API-only, so it isn't a self-hosting option alongside the models here. Xiaomi's MiMo-V2.5-Pro is a 1.02T MoE coding model with ~57.2% SWE-bench Pro and MIT licensing, released in July 2026, requiring 8x H200 or B200 and targeting long-horizon agentic coding tasks with a 1M context window. For teams specifically weighing reasoning quality against API costs, Arcee's Trinity-Large-Thinking (400B total, 13B active, Apache 2.0) is another data point worth adding to this comparison, since the guide to deploying Trinity-Large-Thinking walks through the breakeven math against Claude Opus 4.6 specifically. xAI's Grok 2.5, a 270B MoE model, is a heavier self-hosting commitment than any of these at 8x40GB+ VRAM; the Grok 2.5 deployment guide covers the TP=8 setup and license terms if you're weighing it against this group.

GPT-OSS 120B, GLM-5.1, and DeepSeek V4 are all available on Spheron with on-demand and spot pricing. Pick a single H100 for GPT-OSS 120B, scale up to 8x H200 for GLM-5.1, or run DeepSeek V4-Pro on 8x H200 (V4-Flash fits 2x H200), all with per-minute billing.

Spheron H100 → | On-demand H200 → | View all GPU pricing →

Get started on Spheron →