Wafer reports GLM-5.2 hits 2,626 tok/s on MI355X

• wafer_ai's Hacker News screenshot put the Wafer post at #2 on Hacker News after GLM-5.2 hit 2,626 tok/s/node on AMD MI355X, with the thread claiming roughly 80% of B200 throughput at more than 2x lower cost.
• wafer_ai's write-up reply points to the implementation details: 20k input/1k output, 60% cache hit rate, 2.4 RPS, a ≤5s TTFT knee, and 213 tok/s single stream on a 10k input/1.5k output workload.
• togethercompute's DeepSWE thread framed GLM-5.2 as roughly 80% of Sonnet 5 capability at roughly 20% of the price for software engineering tasks.
• composio's tool-calling benchmark reported 40/41 agentic API tasks completed, while kilocode's code-review test found the model could swing from strong to brittle depending on prompt framing and codebase complexity.
• scaling01's benchmark critique argued public evals are inviting benchmark hillclimbing, and scaling01's MRCR thread pointed to long-context reasoning as the gap that still reopens.

Wafer's write-up buries the best systems details: a 20k in/1k out serving workload, a 60% cache-hit assumption, a ≤5s TTFT knee, FP4 MoE prefill trouble, and single-node scope. A DSpark speculator screenshot shows a separate GLM-5.2-FP8 draft model with 5 layers, block_size=8, a 154,880-token vocabulary, and validation on NVIDIA B300. Kilo Code's blog found prompt wording moved GLM-5.2 code-review results more than extra thinking on a harder codebase.

Wafer's headline number was 2,626 tok/s/node on AMD MI355X, plus 213 tok/s single stream. The claim that made engineers click was the price-performance comparison: about 80% of B200 throughput at more than 2x lower cost, according to wafer_ai's launch thread.

Wafer's write-up describes the measured serving point as:

• Workload: 20k input tokens, 1k output tokens.
• Cache behavior: 60% cache hit rate.
• Traffic level: 2.4 requests per second.
• Latency boundary: knee at ≤5s TTFT.
• Single-stream run: 10k input tokens and 1.5k output tokens at 213 tok/s.
• Scope: single-node results, not a multi-node scaling study.

The launch also spread through a cluster of short reposts and replies. wafer_ai's bare link post carried only the write-up URL, wafer_ai's meme follow-up linked the same post, and andykonwinski's link-only post amplified a separate URL into the same conversation.

Wafer's useful detail was not a new model claim. It was the serving path for a large FP4 MoE on AMD hardware.

According to Wafer's write-up, the slow path was prefill on MI355X, where GLM-5.2's FP4 MoE shapes hit fallback behavior. The tuning target was explicit MoE kernel selection for GLM's shape:

• model_dim: 6144.
• moe_inter: 2048.
• experts: E=256.
• routing: topk=8.
• outcome: 2,626 tok/s/node at 2.4 RPS.
• caveat: no heavy custom-kernel stack for this study, but some framework bugs and single-node limits remained.

That makes the post more interesting than a generic AMD-vs-NVIDIA victory lap. The number came from getting an awkward MoE path onto the right kernels.

The provider screenshots turned the blog number into a distribution story. wafer_ai's OpenRouter table showed Wafer Fast at 169 tps and 99.77% uptime for Z.ai: GLM 5.2, ahead of Fireworks Fast at 98 tps and Weights & Biases at 91 tps in that screenshot.

On Vercel AI Gateway, wafer_ai's live-traffic chart showed Wafer at 287 tps over the prior 24 hours, and wafer_ai's later provider screenshot raised the displayed GLM 5.2 Fast number to 301 tps with 1M context and 0.9s latency.

Availability was already moving into coding surfaces. cline's subscription post introduced a $9.99/month plan for discounted access to GLM-5.2 and other open-weight models, FactoryAI's Droid post added GLM 5.2 Fast hosted by Fireworks AI, and _akhaliq's Claude Code note said they had moved daily usage to GLM 5.2 via hf claude.

Together framed GLM-5.2 as the cost-collapse version of a near-frontier coding model: about 80% of Sonnet 5 capability at about 20% of the price in its DeepSWE analysis. togethercompute's slide release tied that claim to a broader inference-engine talk about serving agentic workloads at trillion-token production scale.

Composio tested a different slice: real API tool use.

In composio's 41-task run, GLM-5.2 completed 40 tasks, or 97.6%, across GitHub, Gmail, PostHog, Notion, and other production APIs. The thread's task design matters because each job required chained tool calls, strict instructions, and often exact JSON output.

The concrete cases were more useful than the aggregate score:

• composio's LaunchDarkly task said GLM-5.2 took 37 tool calls, recovered from 4 rejected calls, and created the correct approval request.
• composio's Gmail pagination task said it followed 19 pages of history API results, found 8 matching messages, and labeled only those.
• composio's Zendesk task said it escalated exactly 1 of 3 tickets to PagerDuty when both required conditions held.
• composio's missed GitHub task said GLM-5.2 stopped when GitHub's API reported 18 results as complete, the same failure mode GPT-5.5 hit, while Opus 4.8 kept digging.

The cost shape was not free intelligence. composio's timing note said GLM-5.2 averaged about 100 seconds per task versus 53 seconds for Opus and used the most tokens of the three, while still coming out cheaper per task at GLM pricing.

Kilo Code's controlled review test found a clean split. On a simple TypeScript backend with planted bugs, kilocode's simple-codebase result said GLM-5.2 caught every serious security bug in every run and landed at 13 to 15 findings out of 16.

The harder codebase broke the smooth story. kilocode's prompt-framing result said wording moved results more than extra thinking, and a strict production-PR approval prompt pushed the model toward a security checklist instead of the planted product bugs.

The repeated miss was system-level consistency. kilocode's cross-route note said GLM-5.2 nailed local bugs visible inside one function, but missed rules that only appeared across search, exports, overdue lists, and other route interactions. kilocode's best-run result put its best hard-codebase run at 7/10, one finding behind Opus 4.8 and two behind GPT-5.5.

The backlash centered on public benchmark incentives. scaling01's critique argued GLM-5.2's jump looked like benchmark hillclimbing because public evals lack hidden test sets and labs can optimize one model for one leaderboard at a time.

Other evidence was domain-specific rather than vibes. deredleritt3r's prinzbench result put GLM-5.2 at 30/99 on a legal reasoning benchmark, behind frontier models and also behind Gemini 3 Pro at 35/99 in that chart. theo's leaderboard post mocked a reasoning leaderboard that placed Sonnet 5, GLM-5.2, and Nemotron 3 Ultra above Fable 5.

The sharper economic framing came from Vtrivedy10's cost-per-solved-problem post, which argued that cheaper open models can approach frontier performance by spending more thinking tokens, but the operational metric becomes cost per solved problem rather than cost per token.

scaling01's MRCR thread argued that long-context reasoning is where the open-model gap reappears. The cited MRCR chart showed GLM's highlighted curve falling hard as context length increased, while the thread claimed Opus 4.8 and GPT-5.5 at 100k+ context matched GLM-5.2 around 16k.

The same thread tied that to two scaling axes:

• RL-training compute for long-context behavior.
• Test-time compute for long-context behavior.

scaling01's long-context follow-up reduced the point to a scaling-law claim: long-context performance is not just a context-window checkbox.

The training-side thread connected GLM-5.2 to a wider shift in agentic RL: long trajectories make value-free credit assignment brittle.

ZhihuFrontier's roundup listed four concrete mechanisms:

• GLM-5.2 moved back to critic-based PPO after trajectory compaction made GRPO's group-wise relative ranking unreliable across uneven sub-trajectories.
• GLM-5.2 used two-stage anti-cheating, rule detection plus an LLM judge, and blocked violating actions without terminating the whole trajectory.
• Qwen paired an Agentic Quality Judge with a Trajectory Monitor for SWE tasks, with the thread reporting cheating rate dropping from 28.57% to 0.56%.
• Qwen's frontend verifier combined a 25.9-item rubric with an interactive judge that clicks, navigates, and tests the UI.
• Qwen's real-world-task work used Span-KTO from implicit user feedback, with negative feedback described as 81.8% high-confidence quality signal.
• GenAC replaced scalar-value prediction with a generative critic that thinks before outputting an integer value, with AIME24 and GPQA value-estimation error dropping by more than half in the roundup.
• OPID mined trajectory-level and step-level skills, routed them by state, matched full-data GRPO with 60% data, beat it with 80% data, and performed worse when both skill types were simply stacked.

That is the technical spine behind the noisy launch week: GLM-5.2's serving economics, tool-use wins, and benchmark disputes all sit on top of a post-training race over long-horizon credit assignment and verifier design.