abliteratedLicense: MIT LicenseVerified 94/100

Uncensored GLM-5.3 Flagship (Abliterated)

zai-org/glm-5.3-abliterated

Full-size GLM-5.3 abliterated flagship model (~85 tok/s, 1M context) engineered for frontier coding, deep security vulnerability analysis, and multi-step Claude Code & Codex sessions.

Input / 1M

$1.40

Cached / 1M

$0.30

Output / 1M

$4.40

Context Window

1,048,576 tok

Max Output

131,072 tok

6-Vector Behavioral Verification Scorecard

Continuous automated probes on bf16.ai screen out over-quantized or broken upstream weights before routing production agent traffic.

Composite Verification Index94 / 100

Math & Symbolic Reasoning Retention86 / 100

Low-Refusal Execution Consistency100 / 100

1M Context & JSON Tool Schema Fidelity100 / 100

P50 TTFT

1895 ms

P95 TTFT

32264 ms

Median TPS

90 tok/s

Zero-Prompt Retention & Data Governance

bf16.ai Prompt Storage: 0 Bytes (Stateless Stream)
Upstream ZDR Retention: 0 Days
Transfer Safeguard: uk_idta_scc
Tokenizer Family: glm4

Quickstart cURL (https://api.bf16.ai)

curl https://api.bf16.ai/v1/chat/completions \
  -H "Authorization: Bearer $BF16_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zai-org/glm-5.3-abliterated",
    "messages": [{"role": "user", "content": "Review this auth middleware diff for session fixation bugs."}],
    "temperature": 0.2
  }'

Recent Behavioral Verification Runs

Suite	Composite	Math	Low-Refusal	Schema	Verified At
bf16-verify-v1	94/100	86	100	100	2026-11-12T18:26Z