Throughput (tokens / sec)
median 140.8tokens/s
Multimodal coding agents with 1M context
Run instantly. Pay only for what you use.
GLM 5.3 Flash is Z.ai's first natively multimodal GLM 5 model, released under MIT: a 320B-parameter MoE that activates just 18B per token, with a 1M-token context window, up to 128K output tokens, and image and video input alongside text.
Its hybrid attention — linear KDA layers plus NoPE sparse MLA — runs on roughly 3× less attention compute with a 4.4× smaller KV cache than the flagship GLM 5.3, so whole-repo prompts, long video transcripts, and heavily cached agent context stay fast and cheap at full 1M context.
The result is GLM 5.2-class agentic work at a fraction of the cost: 84.3 on Terminal-Bench 2.1 and 63.4 on DeepSWE 1.1 at maximum reasoning effort — up from 46.2 for GLM 5.2 — with strict function calling and JSON output in every mode.
18Bactive of 320B parameters
Sparse routing keeps per-token compute near a small model while the full expert pool stays available.
84.3on Terminal-Bench 2.1
Within 0.7 points of Claude Opus 4.8 on end-to-end terminal tasks — from MIT-licensed open weights.
97%avg. cache hit rate on Entrim
Measured on coding-agent workloads — repeated context bills at the cache-read rate.
GLM 5.3 Flash is strongest where agents run long, look at screens, and have to stay cheap — coding and automation loops that mix code, screenshots, and recordings in one context.
Trained on a 30T-token multimodal corpus, it reads screenshots, diagrams, and screen recordings directly — no separate vision model in the loop.
Linear KDA layers plus NoPE sparse MLA cut attention compute by about 3× and the KV cache by 4.4× versus the flagship GLM 5.3 — whole-repo prompts run at full context.
Low, high, and max efforts set thinking depth per request — deep reasoning for the hard steps, shorter traces where a request does not need them.
Function calling and structured JSON output in every mode, with 48.8 on AutomationBench — nearly double GLM 5.2 — for multi-step tool use.
Long-running coding agents
Plans, edits, and verifies across long tool-call chains — 84.3 on Terminal-Bench 2.1 at maximum reasoning effort.
Agents grounded in screenshots and video
Native image and video input means UI screenshots, diagrams, and screen recordings go straight into the prompt without a separate vision model.
Whole-repo and long-document analysis
The 1M-token window holds a full repository or document set in one prompt, and the 4.4× smaller KV cache keeps it affordable.
Browser and workflow automation
48.8 on AutomationBench — nearly double GLM 5.2 — for multi-step tool use across real applications.
Latency-critical short replies
Thinking cannot be switched off — even low reasoning effort spends tokens before the reply. Use DeepSeek V4 Flash in non-think mode for fast conversational turns.
Tight output-token budgets
Reasoning traces run long — about 1.3× the median output tokens in Artificial Analysis testing, and thinking cannot be disabled. Use low reasoning effort for cost-sensitive replies.
Audio input or speech understanding
Inputs are text, image, and video only. Transcribe audio upstream before sending it to the model.
Turbo is Entrim's NVFP4 serving mode for GLM 5.3 Flash — the same open weights, quantized only where it is safe, for higher throughput at a lower price.
Only the routed-expert weights — about 97% of the model’s parameters — are stored in NVFP4. Attention layers, routers, embeddings, the vision tower, and all activations stay in higher precision, so 4-bit noise never touches the layers most sensitive to it.
That cuts the weight footprint by about 70%, which is what makes the 1M-token context practical to serve at Flash pricing.
BF16 vs NVFP4, measured by NVIDIA on GLM 5.3 Flash
| Benchmark | BF16 | NVFP4 | Δ |
|---|---|---|---|
| GPQA Diamond | 92.17 | 92.11 | −0.06 |
| SciCode | 56.21 | 57.69 | +1.48 |
| MMMU Pro | 76.88 | 76.30 | −0.58 |
| AA-LCR (long context) | 71.00 | 71.06 | +0.06 |
| IFBench | 61.30 | 60.54 | −0.76 |
| Terminal Bench 2.1 | 82.58 | 83.15 | +0.57 |
Measured by NVIDIA on its GLM 5.3 Flash NVFP4 checkpoint at temperature 1.0, top-p 0.95. Entrim’s own evals will replace this table when they land.
The table above is NVIDIA's measurement, not Entrim's — validate outputs against your own evals before moving production traffic.
Compare Entrim’s pricing, powered by an optimized inference runtime, against other providers using the same token counts per request.
median 140.8tokens/s
median 1,779ms
84.3Terminal-Bench 2.1
Agentic coding and shell tasks run end-to-end in a real terminal.
63.4DeepSWE 1.1
Multi-step software engineering fixes — up from 46.2 for GLM 5.2.
48.8AutomationBench
Tool-driven automation across real applications and workflows.
Code examples
curl -X POST "https://api.entrim.ai/v1/chat/completions" \ -H "Content-Type: application/json" \ -H "Authorization: Bearer $ENTRIM_API_KEY" \ -d '{ "model": "zai-org/GLM-5.3-Flash", "messages": [ { "role": "user", "content": "Why do they call it a building if it is already built?" } ] }'SDKs & docs
Use familiar SDK patterns with quickstart examples for common production setups.
OpenAI-compatible
Keep the request format your team already knows.
No cold starts
Production requests are served without model spin-up delays.
High-throughput agentic work with 1M context
Dense multimodal coding with thinking control