Home›Companies›Morph
Morph

Fast Models Optimized for Coding Agents

Specialized inference, built for agent workloads Fast Open-Source models: Kimi k3, Deepseek v4 flash, qwen 397b, at 200+ tps Specialized models for applying edits, code search, compaction, and model routing
Active Founders
inference thats optimized for agents
Company Launches
Morph - Abliterated DeepSeek V4.1 Flash at 500 tokens per second
See original launch post

tl;dr Morph now serves DeepSeek V4.1 Flash 1017 with its refusals removed, at 500 tokens per second, so security teams can red team their own product with the same kind of model attackers already use. Access is gated to verified companies. Request access.

The problem

In July, AI agents broke into Hugging Face. When Hugging Face's security team pasted the attack logs into a frontier model to investigate, the model refused. The logs held real exploit payloads, and it couldn't tell an incident responder from an attacker.

The attacker ran with no refusals. The defender got refused.

Attackers have thousands of uncensored checkpoints a search away, and they run them against your infrastructure 24/7. The people who get told no are the ones following the rules: red teams, incident responders, malware analysts, and the security engineer at a company too small for a lab's vetted program.

What we built

DeepSeek V4.1 Flash Abliterated, served on Morph.

  • The DeepSeek V4.1 Flash 1017 checkpoint with the refusal direction removed. It writes the exploit instead of lecturing you.
  • 500 tokens per second. Red teaming is a swarm workload, thousands of long agent loops at once. Speed decides how much of your surface gets tested in an hour.
  • 1M token context. A full incident's logs, a binary's disassembly or a whole repository fits in one request.
  • $0.30 input, $0.03 cached, $1.20 output per million tokens.
  • OpenAI-compatible API. Model id morph-dsv41flash-abliterated.

It has gotten privileged access to sensitive information at 80% of the Series A+ startups we've run it on.

Why it's gated

One uncensored model on one node is not the threat. Anyone can download a checkpoint. A swarm is the threat, and a swarm is capped by compute. We sell that compute, so some of the responsibility is ours.

You request access with a business email, tell us who needs it and what it's for, and we review every request by hand. The full argument: Why We Serve an Uncensored DeepSeek.

Persistent red teaming

Most teams struggle to use the model well. The harness is most of the work. So we also run it for you: our harness, pointed at your app, your API and your infrastructure, nonstop.

  • Reports on everything it finds.
  • Low severity findings meshed into real chains. Three small bugs in the right order is how someone gets in.
  • API access to the model for your own team.

It starts at $6,000 a month with unlimited token usage. If we don't find a high severity vulnerability, we refund you. We're capacity constrained right now, so slots are limited.

How we got here

Morph is inference built for agents. We serve open models for coding agents at high speed. When we pointed an abliterated model at our own infrastructure, it found many vulnerabilities. We fixed them, and then decided every defender should be able to do the same.

Our ask

  • If you run security at a company, request access.
  • If you'd rather we run it against your product, email tejas@morphllm.com.
  • If you know a security team that keeps getting refused by Claude or GPT, send them this.
Previous Launches
Live on OpenRouter and Vercel's AI Gateway
Most Agent failures are behavioral. Catch them at scale
33k tok/sec context compaction for coding agents
Code search limits coding capabiltiies. Moving it out to a specialized model wins.
50% faster than Cognition's SWE-grep - improve Claude Code or Codex via MCP
Instantly apply AI output into code and files (5000+ tok/sec). Build your own coding agents.
Morph
Founded:2025
Batch:Summer 2023
Team Size:2
Status:
Active
Location:San Francisco
Primary Partner:Tom Blomfield