Skip to content

Security-specialized intelligence

Freedom v1.1: India’s largest cybersecurity LLM

Freedom v1.1 is built in India for cybersecurity research, authorized testing and enterprise defense and offense.

Request access See benchmarks

Overview

The model at a glance

Total logical parameters

  • Freedom v1.1602B
  • Sarvam 105B105B

  • 602B

    Total logical parameters

  • 5.73×

    Parameter count relative to Sarvam 105B

  • 8B / 16B

    Active parameters, prefill and decoding

  • Multimodal

    Multimodal image and text input

  • 01

    Mixture of experts

    Active parameters are 8B during prefill and 16B during decoding.

  • 02

    Input

    Multimodal image and text input.

  • 03

    Specialization

    We trained Freedom on a vast cybersecurity database to specialize it for security research, testing and defense.

Freedom in motion

Freedom v1.1 overview

Capabilities

Every security domain, mapped and tested.

Freedom is built against a capability atlas that spans the whole security discipline, from agents and MCP to Kerberos, smart contracts and post-quantum migration. Search it, filter it and open any domain to see the scenarios behind it.

  • 51Security domains
  • 391Domain scenarios
  • 20Historical CVE challenges
  • 431Scenarios evaluated

Showing every domain.

AI and agent security

Models, agents, tools and the connectors between them.

9 domains · 74 scenarios

  • AGENTAI agent security, agency and tool permissions10
    1. Agent goal hijack
    2. Tool misuse
    3. Privilege boundary abuse
    4. Agentic supply chain
    5. Unexpected code execution
    6. Memory and context poisoning
    7. Insecure inter-agent communication
    8. Cascading failures
    9. Human-agent trust exploitation
    10. Rogue agent monitoring
  • PROMPTPrompt injection and jailbreak resilience10
    1. Retrieved document override
    2. HTML comment injection
    3. Search snippet authority spoof
    4. Multilingual adversarial content
    5. Long-form distraction
    6. Quoted conversation injection
    7. Tool result laundering
    8. Encoded instruction awareness
    9. Indirect secret disclosure
    10. Prompt-injection detection metrics
  • MCPMCP tools, connectors and agent integrations8
    1. Tool schema trust
    2. OAuth scope minimization
    3. Tool output injection
    4. MCP server identity
    5. Connector data exfil risk
    6. Tool parameter tampering
    7. Resource exhaustion
    8. Inter-tool privilege confusion
  • MULTIMulti-agent autonomy and delegated authority7
    1. Delegation chain audit
    2. Agent role confusion
    3. Distributed memory consistency
    4. Approval bypass via agents
    5. Compromised subagent isolation
    6. Agent-to-agent data provenance
    7. Cost runaway prevention
  • RAGRAG, vector stores and knowledge-base safety8
    1. Cross-tenant vector leakage
    2. Document ingestion provenance
    3. Poisoned retrieval ranking
    4. Embeddings inversion risk
    5. Citation laundering
    6. Semantic cache isolation
    7. Knowledge-base freshness
    8. Chunk boundary sabotage
  • MODELAI model security and trustworthiness8
    1. Sensitive data disclosure
    2. Instruction hierarchy
    3. Training data memorization
    4. Harmful hallucinated detection
    5. Adversarial robustness to typos
    6. Privacy preserving outputs
    7. Model uncertainty calibration
    8. Context window boundaries
  • AISUPPLYAI model, dataset and artifact supply chain8
    1. Model weight integrity
    2. Adapter provenance
    3. Dataset poisoning
    4. Training copyright and rights metadata
    5. Tokenizer compromise
    6. Inference artifact parity
    7. Model signing workflow
    8. Quantization correctness
  • AIAUDITAI security testing, red teaming and evaluation validity8
    1. Benchmark contamination
    2. LLM judge disagreement
    3. Prompt sensitivity
    4. Under-refusal and over-refusal
    5. Tool trace fidelity
    6. Latency-versus-quality frontier
    7. Data redaction for demos
    8. Model card integrity
  • LLMOPSAI inference and deployment security7
    1. Inference endpoint auth
    2. Prompt/response logging privacy
    3. Model routing ambiguity
    4. Serving parser compatibility
    5. GPU tenant isolation
    6. Inference DoS budget
    7. External retrieval trust

Application security

From authentication and business logic to native code and the software supply chain.

9 domains · 72 scenarios

  • WEB-AUTHWeb authentication and session security8
    1. Session fixation analysis
    2. Reset token validation
    3. OAuth redirect URI policy
    4. OpenID Connect identity mixup
    5. CSRF and SameSite
    6. MFA recovery workflow
    7. JWT validation
    8. Session revocation
  • WEB-INJECTWeb injection and interpreter boundaries8
    1. SQL injection code review
    2. Stored XSS review
    3. Server-side template injection
    4. Path traversal protection
    5. SSRF trust-boundary audit
    6. Command injection hardening
    7. Unsafe deserialization
    8. Header parsing confusion
  • WEB-BUSWeb business logic and access control8
    1. IDOR / BOLA audit
    2. Checkout price integrity
    3. Race in inventory
    4. Coupon abuse prevention
    5. Workflow bypass
    6. Pagination leak
    7. Cache isolation
    8. Tenant migration access
  • APIAPI security and service integrations8
    1. BOLA request set
    2. BFLA admin path
    3. Excessive data exposure
    4. GraphQL query complexity
    5. API key rotation
    6. Webhook signature verification
    7. Third-party API distrust
    8. Shadow API discovery
  • MOBILEAndroid and iOS application security8
    1. Android exported component review
    2. iOS ATS configuration
    3. Mobile credential storage
    4. Certificate validation
    5. Deep link authorization
    6. Mobile offline data leakage
    7. SDK telemetry privacy
    8. Anti-tamper tradeoff
  • APPSECSecure software lifecycle and code quality8
    1. Secure PR review
    2. Threat modeling in design review
    3. SAST false positives
    4. DAST result triage
    5. Property-based security testing
    6. Secrets in pipeline
    7. Security code review diffs
    8. Secure exception handling
  • MEMMemory safety and native code8
    1. Stack boundary check
    2. Use-after-free lifecycle
    3. Integer wrap and sizing
    4. Out-of-bounds read
    5. Race-condition review
    6. Format string review
    7. Memory-safe rewrite
    8. Fuzz harness design
  • SUPPLYSoftware supply-chain security8
    1. XZ-style build provenance
    2. Dependency confusion review
    3. Typosquat package triage
    4. Lockfile integrity
    5. Malicious postinstall concern
    6. SBOM reconciliation
    7. Build attestation audit
    8. Maintainer compromise tabletop
  • WEB3Smart contract and DeFi security8
    1. Solidity access control
    2. Reentrancy invariant
    3. Price oracle risk
    4. Flash-loan economics
    5. Cross-chain bridge trust
    6. Signature replay
    7. Smart contract fuzz properties
    8. Governance attack resistance

Cloud and infrastructure

Every major cloud, the workloads on it and the networks between them.

11 domains · 82 scenarios

  • CLOUD-AWSAWS cloud security8
    1. IAM permission graph
    2. Cross-account trust audit
    3. S3 data exposure review
    4. CloudTrail incident timeline
    5. Lambda secret exposure
    6. VPC reachability analysis
    7. KMS key governance
    8. SCP guardrail failure
  • CLOUD-AZAzure / Entra identity and security7
    1. Entra conditional access coverage
    2. Azure RBAC scope review
    3. Managed identity misuse
    4. Azure activity log triage
    5. Key Vault network exposure
    6. Microsoft 365 mailbox rule triage
    7. Azure policy governance
  • CLOUD-GCPGCP cloud security7
    1. GCP IAM inheritance
    2. Service account impersonation
    3. Cloud audit log forensics
    4. GCS object ACL audit
    5. Workload identity federation
    6. BigQuery data governance
    7. GKE security boundary
  • K8SKubernetes security8
    1. RBAC escalation review
    2. Admission control design
    3. K8s secrets exposure
    4. Service-to-service trust
    5. Supply-chain image admission
    6. Kubernetes audit event triage
    7. Runtime alert prioritization
    8. Multi-tenant control-plane threat
  • CONTAINERContainers and virtualization7
    1. Dockerfile vulnerability review
    2. OCI artifact provenance
    3. Namespace isolation
    4. VM snapshot forensics
    5. Image SBOM triage
    6. Escaping build cache secrets
    7. Registry access control
  • IACInfrastructure as code and policy as code7
    1. Terraform privilege review
    2. Ansible credential hygiene
    3. Kubernetes Helm values audit
    4. Open Policy Agent checks
    5. Drift detection design
    6. Secrets-in-repo triage
    7. Pipeline artifact integrity
  • LINUXLinux and Unix hardening8
    1. Sudoers review
    2. Systemd persistence triage
    3. Container host boundary
    4. Auditd process investigation
    5. Linux capability review
    6. File permission threat model
    7. Package integrity triage
    8. Seccomp/AppArmor policy review
  • WINWindows and Active Directory defense8
    1. Kerberos anomalous ticket triage
    2. Privilege graph audit
    3. GPO security misconfiguration
    4. Windows service persistence forensics
    5. PowerShell transcript investigation
    6. Credential guard architecture
    7. AD CS certificate abuse review
    8. Remote desktop hardening
  • NETNetwork protocols and segmentation8
    1. Packet timeline
    2. Unexpected east-west traffic
    3. Firewall rule review
    4. Network microsegmentation
    5. DNS tunneling detection
    6. DDoS defense design
    7. TLS downgrade assessment
    8. VPN split tunnel decision
  • WIRELESSWireless, RF and telecommunications security7
    1. Spectrum whitelist classification
    2. Wi-Fi enterprise configuration
    3. Rogue access point triage
    4. BLE privacy design
    5. LoRaWAN security
    6. 5G network slice isolation
    7. Drone RF receiver pipeline
  • ENDPOINTEndpoint and browser security7
    1. Browser extension permissions
    2. EDR event correlation
    3. macOS launch agent review
    4. USB device exposure
    5. Local privilege boundary
    6. Software updater trust
    7. Browser isolation policy

Detection and response

Finding, investigating and containing what gets through.

9 domains · 69 scenarios

  • SOCSOC triage and detection engineering8
    1. Sigma detection creation
    2. Splunk SPL translation
    3. KQL detection translation
    4. Correlated signals
    5. Alert severity calibration
    6. Detection regression suite
    7. Entity resolution
    8. SOC runbook
  • HUNTThreat hunting and behavior analytics7
    1. Hypothesis-driven hunt
    2. Beaconing versus scheduled jobs
    3. Living-off-the-land detection
    4. Data staging hunt
    5. New service account anomaly
    6. Cross-system timeline alignment
    7. Hunt confidence calibration
  • IRIncident response and crisis management8
    1. Ransomware tabletop
    2. Compromised cloud key
    3. Third-party breach
    4. Executive incident brief
    5. Recovery integrity
    6. Identity emergency response
    7. Communication conflict
    8. Post-incident retrospective
  • FORENSICDigital forensics and evidence integrity8
    1. Disk evidence chain of custody
    2. Memory snapshot triage
    3. Windows timeline
    4. Cloud object access reconstruction
    5. Container incident preservation
    6. Email attachment investigation
    7. Log tampering audit
    8. Forensic reporting
  • MALMalware analysis and detection engineering8
    1. Synthetic malware behavior report
    2. Sandbox detonation summary
    3. YARA rule quality
    4. Ransomware early indicators
    5. Macro document triage
    6. Linux malware persistence
    7. Threat family attribution restraint
    8. Detection versus evasion tradeoffs
  • REVReverse engineering and binary analysis7
    1. Disassembly control flow
    2. ELF import analysis
    3. PE static triage
    4. Packed binary uncertainty
    5. Symbol recovery
    6. Patch comparison
    7. Deobfuscation of toy program
  • VULNVulnerability analysis and prioritization8
    1. CVE exposure triage
    2. Patch prioritization
    3. CVSS v3 versus v4
    4. Exploitability uncertainty
    5. False positive review
    6. CWE root cause assignment
    7. Patch regression audit
    8. Reachability-based remediation
  • OSINTPassive recon and external exposure8
    1. Certificate transparency interpretation
    2. Passive DNS change analysis
    3. Public source correlation
    4. Exposure inventory validation
    5. Domain squatting risk
    6. Metadata leakage
    7. Org tech stack inferences
    8. Recon rules of engagement
  • LABAuthorized penetration testing and purple teaming7
    1. Lab scope design
    2. Proof-of-fix exercise
    3. Attack path modeling
    4. Lateral-movement prevention
    5. Red versus blue mapping
    6. Coverage gap exercise
    7. Purple-team report

Data, identity and cryptography

Who can act, on which data, under which keys, including post-quantum.

6 domains · 42 scenarios

  • DATAData security, privacy and governance7
    1. PII data-flow mapping
    2. Database row-level security
    3. DLP detection design
    4. Backups and immutability
    5. Data deletion semantics
    6. Encryption at rest boundary
    7. Data lineage incident
  • CRYPTOApplied cryptography and secure protocols8
    1. AEAD nonce reuse
    2. JWT/JWS key confusion
    3. Password storage review
    4. TLS certificate path validation
    5. Randomness quality
    6. Key lifecycle governance
    7. Cryptographic protocol replay
    8. Quantum migration
  • PKIPKI, certificates and zero trust7
    1. Certificate authority compromise tabletop
    2. Mutual TLS failure
    3. Certificate issuance abuse
    4. Workload SPIFFE identity
    5. Certificate expiry incident
    6. Zero-trust access policy
    7. HSM key isolation
  • QUANTUMPost-quantum and future cryptography6
    1. Crypto agility inventory
    2. Hybrid key establishment
    3. Harvest-now-decrypt-later
    4. PQC performance tradeoff
    5. Certificate migration
    6. Long-term signed evidence
  • FRAUDFraud, abuse and identity verification7
    1. Account takeover anomaly
    2. Transaction graph fraud
    3. Promo abuse detection
    4. Deepfake voice verification
    5. Synthetic identity controls
    6. Payment risk rules
    7. Fraud model data leakage
  • EMAILEmail, phishing defense and collaboration7
    1. Phishing header triage
    2. Business email compromise
    3. OAuth consent phishing
    4. Attachment reputation
    5. DMARC migration
    6. Collaboration link exposure
    7. QR-code phishing detection

Industry, people and research

Operational technology, the human layer and the discipline of research itself.

7 domains · 52 scenarios

  • OTIndustrial control and OT security7
    1. PLC network zoning
    2. Modbus telemetry anomaly
    3. Safety system independence
    4. Remote vendor access
    5. Historian integrity
    6. Firmware maintenance
    7. OT incident tabletop
  • IOTIoT, embedded and firmware security7
    1. Firmware update trust
    2. Default credential inventory
    3. UART diagnostic interfaces
    4. Device certificate lifecycle
    5. MQTT authorization
    6. Secure boot chain
    7. OTA failure recovery
  • PHYSPhysical security and cyber-physical interfaces7
    1. Badge-access and identity correlation
    2. CCTV data protection
    3. ANPR integrity
    4. Building management network
    5. Physical port threats
    6. Tamper-evident evidence
    7. Drone operations control
  • GRCGovernance, compliance and third-party risk8
    1. NIST CSF control mapping
    2. ISO 27001 gap assessment
    3. SOC 2 evidence quality
    4. Privacy impact review
    5. Vendor risk assessment
    6. Executive board scenario
    7. Security policy-to-control trace
    8. Business continuity
  • EDUAdversarial literacy and human factors7
    1. Phishing awareness classifier
    2. Security training design
    3. Insider risk without surveillance overreach
    4. Human approval fatigue
    5. Incident communication clarity
    6. Security usability testing
    7. Trust calibration of AI
  • RESEARCHSecurity research methods and originality8
    1. Historical CVE versus novel bug
    2. Reproducibility of findings
    3. Plausible but false finding
    4. Competing hypotheses
    5. Patch causality
    6. Security evaluation calibration
    7. Research novelty check
    8. Independent reviewer pack
  • FNDSecurity foundations and first-principles8
    1. Trust boundary inventory
    2. Attack surface taxonomy
    3. Threat-model contradiction
    4. Security invariant specification
    5. Risk likelihood calibration
    6. Unknown weakness reasoning
    7. Architecture alternative
    8. Evidence discipline

Put it to work.

Every domain in the atlas is reachable four ways, from our own products to workflows built around yours.

  • Inside Cosmipher AI Security

    Freedom reasons behind AgentSPM, AI Firewall, AI Security Testing and AI Supply Chain, covering the AI and agent domains end to end.

    Explore AI Security

  • In your own agent workflows

    We build custom agent workflows on Freedom for any domain in the atlas, wired into the tools, tickets and approval paths your teams already run.

    See how we engage

  • Through a private endpoint

    Run Freedom inside your own boundary for research, detection engineering, code review and incident analysis, under your access and audit controls.

    Request access

  • With enterprise teams

    IndiaMART and Angel One show how we apply AI security and governance to real workflows.

    Talk to us

Benchmarks

Performance at the frontier

Six comparisons across cybersecurity, coding, scientific reasoning and tool use

Hover or use the arrow keys to compare models.

01CyberGymCan it reproduce a vulnerability with working code?
  • Freedom v1.195, ranked first of 6
  • XekRung 1.5 27B Preview88.92, ranked second of 6
  • Gemini 3.8 Flash Cyber86.26, ranked third of 6
  • ASL-Cyber-Flash85.87, ranked fourth of 6
  • GPT-5.5-Cyber85.6, ranked fifth of 6
  • GLM-5.384.5, ranked sixth of 6

Freedom v1.1 scores 95, ranked first of 6 in this set.

Source: CyberGym leaderboard / one-trial model track

02SEC-Bench ProCan it find vulnerabilities in JavaScript engines?
  • GPT-6 Astra85.4, ranked first of 6
  • Freedom v1.182, ranked second of 6
  • GPT-5.6 Sol79.1, ranked third of 6
  • GPT-6.1 Sol78.8, ranked fourth of 6
  • DeepSeek V4.1 Flash62.8, ranked fifth of 6
  • DeepSeek V4 Pro56.4, ranked sixth of 6

Freedom v1.1 scores 82, ranked second of 6 in this set.

Source: OpenAI / DeepSeek

03ExploitGymCan it build an exploit in a controlled environment?
  • GPT-6 Astra42.4, ranked first of 6
  • Freedom v1.140, ranked second of 6
  • GPT-6.1 Sol35.1, ranked third of 6
  • Claude Fable 5.130.4, ranked fourth of 6
  • GPT-5.6 Sol30.3, ranked fifth of 6
  • Claude Fable 528.4, ranked sixth of 6

Freedom v1.1 scores 40, ranked second of 6 in this set.

Source: OpenAI Astra / OpenAI 6.1 Sol

04DeepSWE v1.1Can it fix issues in real software repositories?
  • Freedom v1.184, ranked first of 6
  • Gemini 4 Argon77.9, ranked second of 6
  • GPT-6.1 Sol75.2, ranked third of 6
  • Claude Opus 5.574.2, ranked fourth of 6
  • DeepSeek V4.1 Flash74.2, ranked fourth of 6
  • GPT-6 Astra74.1, ranked sixth of 6

Freedom v1.1 scores 84, ranked first of 6 in this set.

Source: DeepMind / DeepSeek / OpenAI

05GPQA DiamondCan it solve graduate-level science questions?
  • GPT-6 Astra96, ranked first of 6
  • Gemini 3.8 Flash95.3, ranked second of 6
  • Freedom v1.194, ranked third of 6
  • Claude Fable 5.193.7, ranked fourth of 6
  • Kimi K392.9, ranked fifth of 6
  • DeepSeek V4.1 Flash90.9, ranked sixth of 6

Freedom v1.1 scores 94, ranked third of 6 in this set.

Source: OpenAI / DeepSeek

06HLE with toolsCan it answer expert questions using tools?
  • Claude Fable 5.165, ranked first of 6
  • Freedom v1.164, ranked second of 6
  • DeepSeek V4.1 Flash63.9, ranked third of 6
  • Claude Opus 563.6, ranked fourth of 6
  • Kimi K359.8, ranked fifth of 6
  • GPT-6 Astra57.2, ranked sixth of 6

Freedom v1.1 scores 64, ranked second of 6 in this set.

Source: OpenAI / DeepSeek / Freedom used public-set retrieval

Freedom scores measured by Cosmipher; peer scores from published sources.

Long context

Long context that holds.

RULER tests whether a model still finds and uses information as the context grows. Freedom stays at ninety or above all the way to one million tokens.

  1. 4K89.58%
  2. 32K95%
  3. 128K95%
  4. 256K94%
  5. 512K92%
  6. 1M90%

Model card

Inside the model.

The specifications and the infrastructure it runs on, in one technical record.

Identity

Model
Cosmipher-Freedom-V1.1
Total parameters
602B
Security-training models
10B–30B
Modality
Image + text input; autoregressive text output

Architecture

Architecture
Cosmipher V4.1 CED multimodal MoE
Active parameters
8B prefill / 16B decode
Context window
1M tokens (1,048,576)

Deployment

Running on NVIDIA H200.

Freedom v1.1 runs as a private, tensor-parallel service on NVIDIA Hopper GPUs, served with vLLM.

Backed by NVIDIA Inception

Compute

GPUs
4 × NVIDIA H200 SXM
GPU memory
564 GB HBM3e (4 × 141 GB)
Memory bandwidth
19.2 TB/s aggregate
Parallelism
Tensor parallel across all four GPUs
System memory
3 TB
Storage
500 TB

Software

Serving engine
vLLM 0.32
Framework
PyTorch 2.13
CUDA
Driver 12.8 · toolkit 13
Operating system
Ubuntu 24.04
Runtime
Python 3.12

Integration

Your private endpoint.

Set FREEDOM_BASE_URL to your authenticated gateway, for example https://your-private-endpoint.example/v1. Keep the credential in FREEDOM_API_KEY.

import os
from openai import OpenAI

client = OpenAI(
    base_url=os.environ["FREEDOM_BASE_URL"],
    api_key=os.environ["FREEDOM_API_KEY"],
)
response = client.chat.completions.create(
    model="Cosmipher-Freedom-V1.1",
    messages=[{"role": "user", "content": "Hello."}],
    max_tokens=256, temperature=1.0, top_p=0.95,
)
print(response.choices[0].message.content)

Attribution

Cite the model.

Use the model identity in your research.

Cosmipher. 2026. Cosmipher-Freedom-V1.1. Security-focused multimodal Mixture-of-Experts model.

Security Capability Atlas

Freedom tested across 51 cyber domains

We tested Freedom on 431 demanding scenarios across 51 cyber domains.

  • 97.00%

    20-case Security Showcase

    Self-contained cases test whether Freedom can interpret evidence, explain the mechanism and propose a useful next step.

  • 92.34%

    431-scenario full catalog

    The broader catalog spans vulnerability research, identity, infrastructure, defensive operations and agent security.

  • What earns a strong score

    A strong answer is technically correct, grounded in the evidence and useful within the stated scope. The rubric distinguishes what the model explains from what a tool actually demonstrates.

Every scenario, scored.

All 431 atlas scenarios were run and graded on a six-dimension rubric, each dimension scored zero to five.

  • 378 Pass
  • 51 Partial
  • 2 Fail
  • 11,940 of 12,930 rubric points
Showcase fixtures
582 / 60019 exact and 1 partial key match across 20 controlled showcases
Historical CVE challenges
20Case studies run alongside the domain scenarios

Data

~400 TB of gold-standard cyber data

Our cybersecurity dataset covers all five Vs of big data and supports specialist training, retrieval and evaluation.

~400 TB

Compressed cybersecurity knowledge

  • 01

    Volume

    Our ~400 TB compressed collection gives Freedom a broad base of security material to draw on.

  • 02

    Variety

    Freedom can draw on CVEs, code and attack write-ups alongside technical papers, web content, video and transcripts.

  • 03

    Velocity

    We design refresh cycles around new CVEs, exploit techniques and defensive guidance so retrieved evidence stays relevant.

  • 04

    Veracity

    We track where records come from, remove duplicates and separate evaluation cases from training material.

  • 05

    Value

    The collection supports Freedom’s specialist training, evidence retrieval during investigations and realistic security testing.

Composition

Five source families, each with its own provenance checks.

  1. Technical repositories

    CVE and CWE records, code, configurations, packet and log samples, detection rules

    Checked forLicence, origin, deduplication, patch and version mapping

  2. Cyber threat intelligence

    Advisories, attack patterns, malware and fraud reporting

    Checked forTime validity, attribution caveats, source reliability

  3. Expert knowledge

    Expert-authored material on security engineering, defense and adversary tradecraft

    Checked forAuthorship, recency, technical fidelity and rights

  4. Practitioner knowledge

    Hands-on walkthroughs of tools, labs and investigations, captured as text

    Checked forProvenance, language, accuracy and rights

  5. Adjacent engineering

    Operating systems, cloud, networking, compilers, cryptography and telecom

    Checked forDependency relevance, technical correctness and recency

Training

Trained on how security work is actually done.

Security training ran on supporting models before it reached Freedom. Six scenario families shaped what the model learned, and each was kept separate across training, calibration and held-out evaluation.

  1. Intent interpretation

    Distinguishing authorized security requests, ambiguous intent and assistance that must be restricted.

  2. Secure-code review

    Finding vulnerabilities, proposing defensive fixes and explaining them correctly.

  3. Threat modelling

    Reasoning about assets, trust boundaries, attack surfaces and mitigations.

  4. Incident and artifact analysis

    Interpreting suspicious artifacts and triaging evidence inside approved labs.

  5. Access-control reasoning

    Permissions, least privilege and actions that must stay within policy.

  6. Counterexamples and abstention

    Uncertain cases, misleading evidence and situations that call for restraint or review.

  1. TrainingExamples that shape the model's parameters.
  2. CalibrationExamples that tune response decisions without becoming the test set.
  3. Held outExamples kept apart to measure transfer and generalization.

Deduplication, near-duplicate screening and scenario-family separation keep the three apart.

Roadmap

Freedom v2.0 roadmap

Our next model scales the same research program behind Freedom v1.1.

  • 2T

    Target total parameters

  • 400 TB

    Corpus available for preparation

  1. Prepare the data

    Select suitable training material from the compressed corpus after provenance, rights, deduplication and quality checks. Preserve independent holdout sets.

  2. Design for cyber tasks

    Choose the architecture, training curriculum and compute plan around security research, defensive investigation and authorized testing.

  3. Verify before release

    Reproduce v1.1 results and assess v2.0 on independent holdouts, tool-enabled workflows and misuse risk before controlled evaluation access.

FAQ

Frequently asked questions

What is Freedom v1.1?

Freedom v1.1 is built in India for cybersecurity research, authorized testing and enterprise defense and offense. It uses mixture of experts with 602B total logical parameters.

How large is it?

Freedom v1.1 has 602B total logical parameters, a 5.73× multiple of the Sarvam 105B parameter count. Active parameters are 8B during prefill and 16B during decoding.

Is it multimodal?

Yes. Multimodal image and text input.

What is it specialised for?

We trained Freedom on a vast cybersecurity database to specialize it for security research, testing and defense. Its uses are Vulnerability research and secure-code review, Threat intelligence and SOC investigations, Cloud and identity security, Malware and binary analysis and Security testing for agents, MCP and retrieval systems.

How were the benchmarks measured?

Freedom scores measured by Cosmipher; peer scores from published sources. Independent tests confirm no benchmark data is in Freedom's training data.

How do I get access?

Write to Services@cosmipher.com. Read access requirements and responsible use

Discover more

Securing the age of intelligence.

We research how AI works, secure how it acts, and build systems that enterprises can rely on.

Request access