NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI
PyPI, #3569 most downloaded on PyPI
The official Python library for the fireworks API
Latest stable
1.2.33
Last release 14 days ago
08 Nov 2026
Ships on a steady schedule
a new release about every 2 weeks
Some releases are documented
notes for 34 of the last 60 stable releases
Nothing withdrawn
no release was ever pulled
4 years old
206 releases · first in 2022
Drag across the ruler to compare, or tap a release to open it.
deployments, models, deployed models, deployment shapes: return created_by (the creator's email) on Deployment, Model, DeployedModel and DeploymentSha
Full Changelog: v1.2.32...v1.2.33
created_by (the creator's email) on Deployment, Model, DeployedModel and DeploymentShape. Only members of the resource's own account see it; for other callers it is empty.medium thinking effort without changing vocabulary or image-token IDs.Your coding agent can read these notes before it upgrades. Set up the MCP server →
training-sdk: support all, new_session, and none sampler hotload prefix-cache policies through managed service configuration and low-level hotload API
Full Changelog: v1.2.31...v1.2.32
all, new_session, and none sampler hotload prefix-cache policies through managed service configuration and low-level hotload APIs, default to new_session, and retain legacy boolean inputs with migration warnings. Document that the multi-LoRA proxy accepts these policies as no-ops.fw-ai-external/cookbook and document installing released recipes from fireworks-training-cookbook on PyPI.dpo-jobs: accept evaluation_dataset and eval_auto_carveout on DPO job create for held-out preference evaluation
Full Changelog: v1.2.30...v1.2.31
evaluation_dataset and eval_auto_carveout on DPO job create for held-out preference evaluationusage (exact prompt, completion, and cached prompt token counts) on TITO sidecar chat responses, and stream the final usage chunk when a request sets stream_options.include_usage, so harness auto-compaction sees real context usagetraining-sdk: retain adapter-rejected TITO artifacts through the sidecar's configured private trace store and shared retention limits
Full Changelog: v1.2.29...v1.2.30
training-sdk: let a serverless FiretitanServiceClient sample a base model or another run's saved checkpoints without creating a model and without usin
Full Changelog: v1.2.28...v1.2.29
FiretitanServiceClient sample a base model or another run's saved checkpoints without creating a model and without using a concurrent-training-run slot, via standalone sampler routesFiretitanSamplingClient picklable across worker processes (connection state, shared-storage R3 binding, fixed concurrency limit), and add credential-free sampler handles with get_sampler_handle() / FiretitanSamplingClient.from_sampler_handle(handle, api_key=None)model_path values before any requesttraining-sdk: add FiretitanServiceClient.delete_training_session() and FireworksClient.delete_training_session(name) to close a serverless training se
Full Changelog: v1.2.27...v1.2.28
FiretitanServiceClient.delete_training_session() and FireworksClient.delete_training_session(name) to close a serverless training session and release its concurrent-run quota slot immediately instead of waiting for idle expirytraining-sdk: accept keyword-only images= on FiretitanSamplingClient.sample for preprojected vision prompts, and reject unprojected image chunks so sa
Full Changelog: v1.2.26...v1.2.27
images= on FiretitanSamplingClient.sample for preprojected vision prompts, and reject unprojected image chunks so sampler-owned delimiters are not duplicatedtraining-sdk: accept the built-in REINFORCE loss and expose trainer capabilities for REINFORCE, optional score centering, and shared server policy TIS
Full Changelog: v1.2.25...v1.2.26
training-sdk: overlap forward/backward request uploads across calls while retaining upload limits, retry ordering, and optimizer admission barriers; s
Full Changelog: v1.2.24...v1.2.25
training-sdk: preserve Router Replay fallback when validation errors omit rejected input values
Full Changelog: v1.2.23...v1.2.24
training-sdk: record each completed sampling attempt's client-measured latency and generated-token count in ServerMetrics, so callers can compare clie
Full Changelog: v1.2.22...v1.2.23
ServerMetrics, so callers can compare client and serving time from DeploymentSampler.drain_metrics()DeploymentSampler.drain_metrics() now returns one record per completed or failed attempt: every record carries its 1-based retry_attempt, and failed attempts (including non-retryable HTTP errors, truncated streams and responses the parser rejects) set the new ServerMetrics.errorServerMetrics buffer kept for drain_metrics(); samplers that are never drained no longer grow memory for the life of the runclient_latency and client_ttft from the completion request itself with a monotonic clock, excluding the shared R3 format probe and a pre-generation fallbacktraining-sdk: preserve optional Parquet sampling-support references through deployment sampling, Tinker-style sampled sequences, model inputs, and TIT
Full Changelog: v1.2.21...v1.2.22
training-sdk: send projection_head_dim as first-class trainer configuration so any compatible validated training shape can create the requested head,
Full Changelog: v1.2.20...v1.2.21
projection_head_dim as first-class trainer configuration so any compatible validated training shape can create the requested head, while preserving legacy extra_argstraining-sdk: allow TITOSidecar(bind_address=...) to listen on a specific host interface, including Docker bridge gateways; loopback remains the defau
Full Changelog: v1.2.19...v1.2.20
TITOSidecar(bind_address=...) to listen on a specific host interface, including Docker bridge gateways; loopback remains the defaulttraining-sdk: automatically select RDMA weight sync only for dedicated full-parameter trainers and inference replicas that all advertise support; olde
Full Changelog: v1.2.18...v1.2.19
fireworks.lib.dataset_upload.upload_dataset_shards to upload a .jsonl file or a directory of top-level .jsonl shards as one dataset (single upload-endpoint registration, credential-free signed-URL PUTs, then validate_upload), plus collect_dataset_shards and count_dataset_examples helpers; fails closed for CMEK datasetstraining-sdk: add forward-only reads for per-token projection outputs
Full Changelog: v1.2.17...v1.2.18
supports_router_replay on training clients, reported by the trainer at create_model, so callers can decide Router Replay without fetching the base model (None when the trainer does not report it)ModelDetailsUnavailableError (a RuntimeError subclass carrying status_code) from FireworksClient.model_is_moe when the model record cannot be read, so callers can tell an inaccessible model (403/404) from a control-plane failuretraining-sdk: add lora_init_method to create_training_client (init_method on create_lora_training_client) to pick each model's LoRA initialization at
Full Changelog: v1.2.16...v1.2.17
lora_init_method to create_training_client (init_method on create_lora_training_client) to pick each model's LoRA initialization at create time: "kaiming" (default) or "nora" (NoRA-init)training-sdk: add projection_head_dim to managed service and training-client creation for train-only projection heads, and output="projection" for for
Full Changelog: v1.2.15...v1.2.16
projection_head_dim to managed service and training-client creation for train-only projection heads, and output="projection" for forward_backward_custom gradients with respect to projection outputstraining-sdk: retain per-token top-logprob token IDs and values in sampling results and TITO trajectory artifacts
Full Changelog: v1.2.14...v1.2.15
training-sdk: select Parquet R3 using a small completion request; older inference servers retain inline routing without a capability endpoint
Full Changelog: v1.2.13...v1.2.14
r3_ttl_seconds for shared Parquet routing files; the default retention is six hoursdeployments: add accept_shapeless_risk opt-out on deployments.create for intentionally creating a deployment without a pre-validated shape; document t
Full Changelog: v1.2.12...v1.2.13
accept_shapeless_risk opt-out on deployments.create for intentionally creating a deployment without a pre-validated shape; document the deployment_shape="default" sentinel that asks the server to pick a default validated shapeaccept_shapeless_risk to DeploymentConfig for the manual shapeless creation path, rejected client-side when combined with deployment_shapeturn/model_tokens, turn/requested_output_tokens, trajectory/model_tokens_processed, agent/wall_seconds, and the debug/* artifact-write family) and remove the observe_agent_wall API; trajectory summaries and Harbor trial phase brackets carry the same timing informationtraining-sdk: deprecate explicit reference shapes for LoRA policies in favor of shared-session base references
Full Changelog: v1.2.11...v1.2.12
comms="v2"; missing or unknown capabilities preserve comms="v1" and its original request format without a manual flagtraining-sdk: add environment-local TITO sidecars with full-history prompt construction by default and an experimental incremental renderer protocol
Full Changelog: v1.2.10...v1.2.11
training-sdk: expose merged-base output precision with source-format export as the default
Full Changelog: v1.2.9...v1.2.10
training-sdk: preserve per-request affinity for serverless training sampling
Full Changelog: v1.2.8...v1.2.9
training: default managed SFT/DPO and dedicated Training API trainers to reservation-first placement with an explicit opt-out
Full Changelog: v1.2.7...v1.2.8
training: support reservation-first placement for managed SFT/DPO and dedicated Training API trainers
Full Changelog: v1.2.6...v1.2.7
training: support CMEK-encrypted managed RFT sampler hot-loads and isolate reference trainers
Full Changelog: v1.2.5...v1.2.6
training-sdk: log backend training shape override warnings
Full Changelog: v1.2.4...v1.2.5
training-sdk: expose MoE model capability lookup for Router Replay
Full Changelog: v1.2.3...v1.2.4
require generated promotion provenance before public SDK releases
Full Changelog: v1.2.2...v1.2.3
## 1.2.2 (2026-07-21) Full Changelog: v1.2.1...v1.2.2
Full Changelog: v1.2.1...v1.2.2
serverless-rl: single retry owner + structured sampling error (SDK)
Full Changelog: v1.2.0...v1.2.1
## 1.2.0 (2026-07-11) Full Changelog: v1.2.0-alpha.89...v1.2.0
Full Changelog: v1.2.0-alpha.89...v1.2.0
training-sdk: sanitize public package contents
Full Changelog: v1.2.0-alpha.88...v1.2.0-alpha.89
training-sdk: default sampler concurrency to adaptive
Full Changelog: v1.2.0-alpha.87...v1.2.0-alpha.88
training-sdk: declare urllib3 runtime dependency
managed: keep auto-shape selection when trainer.extra_args is set
training: propagate auto-shape max context metadata
Full Changelog: v1.2.0-alpha.84...v1.2.0-alpha.85
training-sdk: surface serverless session (on the service) + run (on the training client)
Full Changelog: v1.2.0-alpha.83...v1.2.0-alpha.84
public-repos: make SDK pyproject the version source
Full Changelog: v1.2.0-alpha.82...v1.2.0-alpha.83
cookbook/rl: expose lora_alpha in rl_loop and managed provisioning
Full Changelog: v1.2.0-alpha.81...v1.2.0-alpha.82
training-sdk: full-distribution RL rollouts + default LoRA alpha=32
Full Changelog: v1.2.0-alpha.80...v1.2.0-alpha.81
rlor: merged-base LoRA save + reject ineffective warm_start_from LoRA
Full Changelog: v1.2.0-alpha.79...v1.2.0-alpha.80
## 1.2.0-alpha.79 (2026-06-09) Full Changelog: v1.2.0-alpha.78...v1.2.0-alpha.79
Full Changelog: v1.2.0-alpha.78...v1.2.0-alpha.79
rlor: centralize rollout cleanup mode
Full Changelog: v1.2.0-alpha.77...v1.2.0-alpha.78
serverless: sft_loop serverless mode + SDK session-checkpoint promote
Full Changelog: v1.2.0-alpha.76...v1.2.0-alpha.77
## 1.2.0-alpha.76 (2026-06-04) Full Changelog: v1.2.0-alpha.75...v1.2.0-alpha.76
Full Changelog: v1.2.0-alpha.75...v1.2.0-alpha.76
## 1.2.0-alpha.75 (2026-06-03) Full Changelog: v1.2.0-alpha.74...v1.2.0-alpha.75
Full Changelog: v1.2.0-alpha.74...v1.2.0-alpha.75
## 1.2.0-alpha.74 (2026-06-01) Full Changelog: v1.2.0-alpha.73...v1.2.0-alpha.74
Full Changelog: v1.2.0-alpha.73...v1.2.0-alpha.74
autorelease single source of truth
Full Changelog: v1.2.0-alpha.72...v1.2.0-alpha.73
Full Changelog: v1.2.0-alpha.73...v1.2.0-alpha.73
bootstrap public SDK release workflows
Full Changelog: v1.2.0-alpha.70...v1.2.0-alpha.71
Full Changelog: v1.2.0-alpha.70...v1.2.0-alpha.71
## 1.2.0-alpha.70 (2026-05-15) Full Changelog: v1.2.0-alpha.69...v1.2.0-alpha.70
Full Changelog: v1.2.0-alpha.69...v1.2.0-alpha.70
sdk/training/hotload: forward snapshot URI via x-fireworks-hot-load-source-url header
Full Changelog: v1.2.0-alpha.68...v1.2.0-alpha.69
## 1.2.0-alpha.68 (2026-05-06) Full Changelog: v1.2.0-alpha.67...v1.2.0-alpha.68
Full Changelog: v1.2.0-alpha.67...v1.2.0-alpha.68
## 1.2.0-alpha.67 (2026-05-05) Full Changelog: v1.2.0-alpha.66...v1.2.0-alpha.67
Full Changelog: v1.2.0-alpha.66...v1.2.0-alpha.67
restore prerelease versioning in release-please-config.json
sdk/deployment: SSE-truncation retry hardening
promote: accept 4-segment name= kwarg
Full Changelog: v1.0.1-alpha.63...v1.1.0-alpha.64
One column per quarter.