Changelog

​

Added

• Local model and embeddings models loading to unlock faster cold starts:
  • --llm-model-path <path>
  • --embeddings-model-path <path>
• KV-cache size controls to control how much of your system resources are allocated to inference generation. Default: --kv-size-mode auto
  • --kv-size-mode <auto|min|medium|max> (apply sizing by mode: auto | 33% | 66% | 100% of available limit)
  • --kv-size-tokens <int> (default and min are 1024, target cache in tokens, fallback to default: --kv-size-mode auto)
  • --kv-size-gb <float> (default is 1.0, target cache in GB. If less than 1, fallback to --kv-size-mode or --kv-size-mode auto)

​

Fixed

• Fixed intent resolution comparison.

​

Added

• Faster, more resilient networking: QUIC transport (opt-in) runs alongside TCP on the same port with automatic TCP fallback; upgraded to QUICv1 for lower latency and smoother handshakes.
• Connect through tough networks: NAT hole-punching (DCUtR) upgrades relayed links to direct connections with retry/backoff and Prometheus metrics for visibility.
• Smarter peer knowledge: PeerStore now tracks multiple addresses per peer to improve reachability.
• Configurable relay mode: Optional GossipSub relay acceptance/validation (off by default). Enable with FT_GOSSIPSUB_RELAY_ENABLED=true.
• Easier bring-up & testing: IPFS bootstrapping option; Docker Compose to run multiple nodes on one machine; Dockerfile now works with Podman.
• Better observability: New Prometheus metrics for DCUtR operations, peer analysis, and connection state.

​

Changed

• Less waste, better throughput: Participation logic tuned so a node skips very slow or ineffective rounds.
• Lower latency by default: Dialer prioritizes QUIC addresses when available.
• Cleaner logs: Error noise reduced where retries are automatic; clearer transport-type logging.
• Smarter routing: Improved address policy for relayed vs. direct connections.
• Safer defaults: Updated node/network configuration; DCUtR is enabled conditionally via config.

​

Added

• Extra validation before capsule auto-start so it doesn’t interfere with app authorization process.
• Compatibility check: if nvidia-smi is missing and the system is not macOS, the app now shows System Incompatible notification.

​

Changed

• GPU-based systems will now always run GPU capsule.
• Simplified onboarding: Go to node button now closes the dialog instead of redirecting to another page.
• (macOS) Improved active window behavior:
  • App window always creates a menu bar entry.
  • When switching to app, a menu bar window is opened.
  • When launching the app when it is already running, its window is opened instead of launching another instance of the app.

​

Fixed

• Notifications:
  • Removed interaction options from notifications that have already been resolved.
  • Fixed globe icon color on notifications — it was invisible in light theme.

​

Added

• Bug reporting: System information is now collected and sent as a separate file (contained in the log archive) when submitting a bug report.
• Added “Documentation” item to the tray menu.
• (macOS) Launch Agent is now used instead of AppleScript for the auto-launch option — now the app should’t ask for additional permissions.

​

Changed

• (macOS) Updated app icons.
• Clarified titles in a few app messages.
• Adjusted memory check — it now uses available VRAM instead of total VRAM.
• VRAM warning is now triggered only if less than 2 GB of free VRAM is available.

​

Fixed

• Notification Center:
  • Some feedback messages no longer appear in the notification center for clarity.
  • Some warnings and critical notifications (e.g. Heavy Load, Connection Failed) will no longer duplicate if an active notification with the same code already exists.
• Fixed Connection Failed notification during onboarding – now only shown at the creation stage when an internet connection is actually needed.
• Fixed model list handling — now filtering out unnecessary models using the automation:false tag, they will no longer appear in automatic mode.
• Improved stability when losing internet connection:
  • When a download is interrupted due to lost connection, the model will no longer be mistakenly marked as downloaded and re-downloaded by the capsule.
  • Better separation of Connection Failed error states during node operation.
  • Reconnection notifications are now more stable both when the capsule is stopped and when it is running. It retries pings and disappears when connection is established; if automatically paused, it resumes automatically.
  • Service connection status no longer disappears after a timeout when offline — it now clears only after a successful ping or when the node connection issue is resolved.
  • Downloads now reliably restart once internet connection is restored (previously, this worked inconsistently).
• Fixed broken dock menu on macOS
• Fixed incorrect behavior of tray options in “Open in window” mode.

​

Changed

• Increased maximum round capacity

​

Added

• Added machine specs logs
• Added kv-cache init log

​

Changed

• Fixed gguf file merging

​

Fixed

• Price calculation when creating a request through rpc API has been fixed

​

Added

• Add --announce flag.

​

Changed

• Removed propagation to swarm and inference manager during createIntent

​

Fixed

• Creating an intent on behalf of another author via RPC API

​

Added

• Added support gpt-oss
• Updated metadata values for symbols

​

Changed

• Updated llama cpp bindings to b6118

​

Added

• Implemented new ranking algorithm
• Added warning message when node TPS is below optimal

​

Fixed

• Fixed gossip performance metrics

​

Added

• Startup timeout (180s) to prevent hangs during initialization.
• Detailed startup logs for Swarm and Blockchain readiness.
• New exit codes:
  • 10: Startup error
  • 11: Startup timeout
• Added delegator address to rpc API

​

Changed

• Extended capsule API for making participation decisions

​

Fixed

• Readiness check loop now handles shutdowns correctly, avoiding hot loops and high CPU load.

​

Added

• Added health check to API
• Added non-blocking metadata request
• Added context length checker to evaluate-participation request
• Added support structured outputs
• Added new ranking schema for v1 ranking
• Implemented Efficient KV Cache Reuse for ranking

​

Added

• Offline mode support for using locally cached models.
• Notification system to confirm successful worker initialization before warm-up.

​

Changed

• Adjusted worker initialization sequence for improved launch stability.

​

Added

• Async block gap processing to detect and fill missing blockchain data automatically.
• Minor stability improvements for consensus submissions.

​

Changed

• Removed relay-specific logic from the network module to simplify architecture.

​

Fixed

• Corrected label assignments in the inference metrics output.

​

Added

• Protocol exit codes for improved error signaling.
• Optimized RPC by implementing raw block polling with bloom filter pre-checks to reduce event overhead.
• Added operational metrics (node state, request tracking).
• Included blockchain-specific metrics.

​

Changed

• Updated console time format for improved readability.
• Reintroduced CLI arguments for backward compatibility with CLI and GUI apps.
• Upgraded Rust version to 1.85.0.

​

Fixed

• Corrected RPC request handling to ensure proper use of swarm consensus.

​

Added

• Graceful shutdown handling for clean termination.
• Defined exit codes for clearer process state signaling and bug reporting.

​

Added

• Web crawler integration.

​

Changed

• Disabled Python worker by default.

​

Added

• Implemented polling-based staggered submission for inference resolutions with position-based delays.
• Added initial version of new ranking algorithm.
• Added fallback RPC support with automatic failover and retry layer for improved blockchain connectivity.

​

Changed

• Changed blockchain event query error logging from error to warning level since query_events is automatically retried.
• Changed swarm connection loss logging from error to warning level when insufficient peers are available.
• Changed capsule binding address from localhost to 0.0.0.0 to allow access from external clients.

​

Changed

• Updated core dependencies: llama.cpp and ft-tools.
• Upgraded Rust version to 1.85.0.

​

Added

• Introduced support for multiple bootstrap nodes to improve peer connectivity.

​

Changed

• Increased blockchain safety buffer and inference submission offset for better transaction stability.

​

Added

• Introduced /metrics endpoint for Prometheus-based performance analysis.
• Enabled new peer discovery method during node initialization.
• Added node startup synchronization.
• Integrated Identify protocol for peer recognition.
• Forwarded sorted ranking results in HTTP responses.

​

Changed

• Renamed 42T token to FOR.
• Updated bootstrap nodes.
• Upgraded Rust toolchain to 1.84.0.

​

Fixed

• Reworked network interaction to resolve sync issues.
• Ignored early requests received before node startup completed.
• Cleaned up log to reduce noise.

​

Added

• Integrated Kademlia protocol for peer discovery during initialization.
• Implemented node startup synchronization to ensure readiness.
• Added Identify protocol support for recognizing peer nodes.
• Forwarded sorted ranking results included in HTTP response.

​

Changed

• Refactored blockchain module to use alloy instead of ethers.

​

Fixed

• Resolved bootstrap node initialization issue.
• Ignored incoming requests created before node startup is complete.

​

Fixed

• Resolved issues with LLM warm-up during capsule initialization.

​

Changed

• Added additional logs for development and debugging purposes.

​

Changed

• Changed default tokenizer used for completions.

​

Added

• Included max_size parameter in the completions request to the capsule for better input control.

​

Added

• Option to include reasoning tokens in output.
• Parameter to adjust reasoning token count.
• max_size field added to completion requests.

​

Changed

• Increased default reserved memory allocation.

​

Added

• max_size field to the /devs request for controlled query sizing.

​

Added

• Embedding truncation for improved memory efficiency.
• Support for using the model’s maximum context size.
• Optimized input token decoding for better inference speed.
• Tokenizer object now returned in /metadata.
• Dynamic selection mechanism for context length.
• GPU offloading support for KV cache.

​

Changed

• Updated development dependencies.

​

Fixed

• Fixed batch prefill logic to improve prompt handling stability.

​

Fixed

• Tokenizer_id in /devs request is now optional, improving support for lightweight queries.

​

Fixed

• Correctly handles empty tokenizer cases in submitted intents.

​

Added

• GET /metadata endpoint for capsules to fetch version, token limits, and tokenizer info.
• Participation now estimates performance limits from capsule metadata.
• New ranking method added that returns a confidence score per candidate.
• Load balancing system introduced to evenly distribute requests across nodes.
• New experimental pairwise ranking algorithm implemented.
• Failsafe added to exit if allowance approval fails at minimum threshold.

​

Added

• Added Qwen3 model support.

​

Changed

• Updated development dependencies.

​

Changed

• Added handling for empty model output in get_completions response.
• Updated system prompt for better context initialization.

​

Changed

• Temporarily disabled Python interpreter.
• Fixed emoji rendering issue in output.
• Added structured input format for evaluate-participation.

​

Added

• Node balance validation before joining a round.
• Auto-approval of 42T token allowance for smoother participation.

​

Added

• New /v2/ranking API endpoint.
• Support for structured input objects.

​

Changed

• Ranking results are now derived directly from logits.

​

Added

• /metadata endpoint for diagnostics and external tooling.

​

Changed

• Replaced Notify/result store with mpsc streaming for improved data flow.

​

Added

• Brave API integration for web search.
• Python execution support.

​

Changed

• Updated Rust version to 1.84.0.

​

Changed

• Increased max duration for inference join attempts.
• Modified participation probability to retain a proportional number of idle nodes for upcoming requests.

​

Changed

• Optimized client-side gas consumption.

​

Fixed

• Patched retry logic covering blockchain context, balances, token approvals, and intent resolution.

​

Added

• Support for custom private Rust crates.
• Token generation speed: 97.63% parity with llama.cpp.
• Performance metrics logged in debug output.

​

Changed

• Codebase refactor for modularity and readability.
• Updated ranking prompt and generation parameters.
• Updated papaya crate to v0.2.1.
• Separated token-to-text from generation worker.

​

Added

• Off-chain consensus layer via P2P.
• Automated deployment pipeline for new releases.

​

Changed

• Refactored message queue to ignore certain request messages and prevent storage overhead.

​

Changed

• Code refactor for maintainability.

​

Fixed

• Runtime resource usage.

​

Added

• New embeddings model.
• Flash attention for faster processing.
• max_tokens support for ranking requests.

​

Changed

• Expanded context window for embeddings. ranking logic.

​

Fixed

• LLM context window bug.
• General resource optimization.

​

Added

• Configurable cache location for uploaded models.

​

Changed

• Upgraded to llama.cpp version b4902.
• Updated Linux CUDA configuration.

​

Changed

• Reduced MAX_PARTICIPATION_AMOUNT from 12 to 7 to lower gas costs.

​

Fixed

• Fixed multithreading desync issues in active round processing.

​

Added

• Added delay before initializing BlockchainBridge.

​

Added

• Retry logic for blockchain sync, balances, approvals, and intents via retry_on_error!.
• New u256_frac_mul! macro for precise U256 math.
• Time sync with on-chain deadlines for accurate state validation.

​

Changed

• Code formatting improvements via rustfmt.toml.
• Dependency Updates
• Upgraded rand crate from 0.8 → 0.9.0.
• Added semver crate at 1.0.0.

​

Fixed

• Optimized mathematical functions in the ranking module.

​

Added

• Implemented forced exit on swarm disconnection to improve network stability.

​

Changed

• Enhanced log messages for better debugging and clarity.

​

Fixed

• Resolved an issue causing failures in swarm reconnection.
• Corrected the ranking algorithm for improved inference accuracy.
• Fixed a bug where ParticipateInInference continued running unexpectedly.