Use moonshotai/Kimi-K3.1-Pro, with text, image, and tool-call support, through get_tokenizer() and get_renderer(). The renderers are kimi_k31_pro_max_reasoning (the default), kimi_k31_pro_high_reasoning, kimi_k31_pro_low_reasoning, kimi_k31_pro_disable_thinking, and kimi_k31_pro_no_reasoning_effort. get_lr() has no calibrated learning rate for it, so set one explicitly.
Pass TrainOnWhat.LAST_ASSISTANT_TURN to the tml_v0 renderer and to tinker_messages_to_sft_message_list() to weight the assistant messages after the last user message. Before, both raised NotImplementedError.
Changes
None.
Fixed
None.
Deprecations
None.
Breaking changes
Direct callers of TmlRenderersTokenizerAdapter(name_or_path) must now supply a tml_tokenizer argument. Prefer get_tokenizer(), for example get_tokenizer("thinkingmachines/Inkling-Small"), to select and construct the correct adapter.
RL and SL trainers report per-step and cumulative token usage and estimated USD cost. Rates are fetched once per process from published pricing; if unavailable, token counts remain available. RL totals include discarded constant-reward groups and KL scoring, but do not represent complete billed usage.
Call tinker_messages_to_openai(...) in third_party.openai_compat to build OpenAI chat inputs from cookbook messages, including tool calls and declarations, reasoning, images, and audio. openai_messages_to_tinker(...) now also converts reasoning, images, and audio.
Convert cookbook messages to and from tml-renderers message lists through third_party.tml_renderers_compat, including supervised weighting and multimodal content.
Use thinkingmachines/Inkling-2-Small with get_tokenizer() and the tml_v0 renderer. LoRA parameter counts are included. get_lr() has no calibrated learning rate for it, so set one explicitly.
BenchmarkConfig.for_model() now returns evaluation defaults for every Inkling model, including the 256K :peft:262144 variants. Before, it raised ValueError for them.
Call tokenizer_utils.encode_without_special_tokens(tokenizer, text) to encode raw text with Hugging Face and tml-renderers tokenizers alike.
Changes
Async RL skips sampler-only weight exports after all training rollout workers finish, while still refreshing weights needed for evaluation, checkpoints, and KL measurements.
Fixed
Preserves original SDFT teacher token positions across masked gaps so teacher and student tokens remain correctly aligned.
Backs up the metrics file and removes rows at or beyond the resumed step before logging again, preventing duplicate or out-of-order rows when resuming an earlier checkpoint.
Completes partial training steps, final evaluation, and environment cleanup during async shutdown.
Returns ParseTermination.MALFORMED with decoded text when the tml_v0 renderer parses an unsupported text-form tool call, instead of failing or producing an empty tool name.
Deprecations
None.
Breaking changes
Async RL now trains completed rollouts regardless of age instead of discarding stale rollouts and retrying their problems. max_steps_off_policy limits sampling ahead to groups_per_batch * (max_steps_off_policy + 1) groups, increasing possible sandbox concurrency. Size environment pools accordingly; use max_steps_off_policy=0 when on-policy execution is required. Observed staleness is reported in async/staleness_min, async/staleness_max, and async/staleness_mean.
Async RL and the math_rl, rubric, and prometheus_experimental recipes now default to StableReinforce (alpha), including its default sampler KL coefficient of 0.01. Set AsyncConfig.loss_fn="importance_sampling", or pass loss_fn=importance_sampling in those recipes, to retain the former objective. Conflicting nondefault Config.loss_fn values in async runs now raise ConfigurationError. Other synchronous RL configs retain their existing importance_sampling default.
StableReinforce requires sampled-token provenance: custom completers must populate TokensWithLogprobs.sequence_id, or callers must choose a different loss. With an active KL penalty, set kl_discount_factor=0 or use another loss; discounted prompt-token advantages are not supported by StableReinforce.
Async RL now keeps evaluation metrics that finish after newer training steps. To do this, async training and late evaluations log independent metric events, so W&B internal steps count metric arrivals rather than training updates. For charts and consumers, use training/step and eval/snapshot/training_step as the model-version axes. Evaluation rows also carry the cost and elapsed-time coordinates of the weights they evaluated.
The cookbook now calls the unified APIFuture methods of tinker 0.38 (sample, forward, and others) instead of the *_async methods. Objects that stand in for a Tinker client must provide these methods; for example, the capture proxy's SamplingClientLike now requires sample. See the migration guide.
rl.metrics.incorporate_kl_penalty(), rl.metrics.compute_post_kl(), and rl.train.prepare_minibatch() now also return token usage. Unpack the extra value: metrics, usage = await incorporate_kl_penalty(...), metrics, usage = await compute_post_kl(...), and data_D, metrics, usage = await prepare_minibatch(...).
Set rl.train.Config.num_epochs to train for multiple passes over an RL dataset. EpochCycledRLDataset supports deterministic epoch cycling while checkpoints and resume retain global step numbering.
Set sandbox_factory on the Harbor distillation CLI config, directly or through an import string, to use a custom sandbox backend.
Set session_name in recipe CLIs and in the DPO and distillation configs to name the Tinker session. By default, a recipe uses its run name or a supplied wandb_name. Recipes without a run name use cookbook-<recipe>, and multi-stage recipes append their stage.
Changes
SL and RL trainers now finish their Tinker sessions when training ends, including successful, failed, and interrupted runs.
RL and SL training now log a warning when max_steps exceeds num_epochs times the dataset length. Training still stops after the last epoch.
Fixed
Finishes rolling checkpoint saves before recording the final checkpoint, preventing resume from pointing to a deleted checkpoint.
Splits held-out source rows before interleaved-dataset oversampling, preventing duplicate holdouts and leakage into training. This changes which rows are held out, and a test_size larger than the sources allow now raises ValueError.
Smaller evaluation fixes: AIME grading rejects fractional and out-of-range answers instead of truncating them, which can lower AIME scores. Resumed evaluations keep their token and turn counts. Deleting an evaluation run also removes unfinished transcripts. Importing the custom-evaluator example no longer loads a tokenizer or calls the service.
Skips None metric values in dict_mean instead of crashing.
Consumes the final queued batches in streaming-minibatch async training before shutdown completes.
Prevents "Already borrowed" tokenizer errors during concurrent Qwen rendering.
Evaluates the final trained weights in synchronous RL loops when evaluation is enabled.
Deprecations
None.
Breaking changes
weights.download() no longer accepts sampler_weights/ checkpoints after removal of the SDK archive API. Export from a training client with save_weights_external() and download the resulting external_weights/ path instead.
The standalone recipes.rl_loop now defaults to StableReinforce (alpha) instead of importance_sampling. Pass loss_fn=importance_sampling to retain the previous objective; loss_fn also supports cispo and gspo.
Dependencies
tinker >=0.34.0 → >=0.35.0
litellm (litellm and all extras) >=1.80.0,!=1.82.7,!=1.82.8 → ==1.105.1
Adds supervised.nll_evaluator.SamplerNLLEvaluator for held-out NLL evaluation through a sampling client. Use .from_dataset(...) and supply a tokenizer to also report bits per byte.
Adds decision-model helpers for scoring labeled choices and training on either a chosen label or a target probability distribution. Use DecisionRenderer, sample_decision, and the datum helpers in recipes.decisions.
Adds a long-context, many-turn counting recipe for checking RL numerics with a paginated document-reading tool.
Use qwen3_5_preserve_thinking with get_renderer() to retain reasoning across Qwen3.5/Qwen3.6 tool-use turns.
Use GLM-5.3-Flash with the glm5_3_max_reasoning (default), glm5_3_high_reasoning, and glm5_3_low_reasoning renderers. Use DeepSeek-V4.1-Flash and DeepSeek-V4.1-Flash-1017 with the new deepseekv4_1, deepseekv4_1_low_reasoning, deepseekv4_1_max_reasoning, deepseekv4_1_disable_thinking, and deepseekv4_1_preserve_thinking renderers. Use Nemotron-3.5-Super-VL, including image input, with the nemotron3_ultra renderers. GLM-5.3-Flash and DeepSeek-V4.1-Flash take text only. LoRA parameter counts are included. get_lr() has no calibrated learning rate for these models, so set one explicitly.
Set session_name on SL and RL training configs to name the Tinker session. When unset, wandb_name is used if provided.
Pass tinker://<run_id>/external_weights/<name> or tinker://<run_id>/weights/<name> paths to weights.download(). A training checkpoint downloads through a temporary encrypted export that is decrypted locally; this starts a training session on the checkpoint's base model. Files are installed only after the complete download succeeds.
Changes
The forecasting recipe adds a GLM-5.3 example and updates defaults to learning rate 1e-4, 128 steps, evaluation every 16 steps, and checkpoints every 64 steps.
SFT and DPO held-out NLL evaluation now uses sampling clients.
Training traces now include external weight exports through TrainingClient.save_weights_external_async.
Fixed
Honors save_every in the guess-number and twenty-questions recipes.
Trims leading and trailing whitespace from list-shaped Qwen3.5 and Qwen3.6 content, as string content already was, so sampled tool-use turns re-render consistently. This changes rendered tokens for such content.
Omits empty header chunks from generation prompts after consecutive tool responses, preventing HTTP 400 errors, including with Nemotron parallel tool calls.
Handles Modal-side sandbox cancellation as SandboxTerminatedError without canceling the entire training step or discarding healthy rollouts. Cancellation of the caller still propagates.
Deprecations
supervised.nll_evaluator.NLLEvaluator is deprecated. Create a sampling client with TrainingClient.save_weights_and_get_sampling_client() and use SamplerNLLEvaluator instead. No removal version or date is announced.
The checkpoint-copy script is deprecated in favor of ServiceClient.copy_weights(). The script retains its behavior; removal is planned for a future release without a specified version or date.
Breaking changes
Recipe outputs now default to ~/tinker-runs instead of /tmp, and re-downloadable caches use ~/.cache/tinker-cookbook. Set TINKER_COOKBOOK_RUNS_DIR to choose another root, or use cli_utils.runs_path() to build a path under it. Explicit output paths work as before. Existing outputs are not moved: re-run data-preparation steps, or pass the old paths, for data created by an earlier release.
W&B logging now also accepts credentials saved by wandb login. A run that sets wandb_project without W&B credentials now raises ConfigurationError instead of continuing without W&B. Run wandb login or set WANDB_API_KEY, set WANDB_MODE=offline, or leave wandb_project unset.
Trains the <|observation|> or <|user|> terminator on every GLM-5.3 assistant turn, rather than only the final turn. Prompt-supplied thinking-prefill tokens carry no loss. This changes loss weights for multi-turn supervised examples.
Use GLM-5.3 with the glm5_3_low_reasoning, glm5_3_high_reasoning, and glm5_3_max_reasoning renderers through get_renderer(), including tool calling. get_lr() has no calibrated learning rate for it, so set one explicitly.
Run the forecasting recipe to train on Prophet Arena binary prediction markets, with a chronological train/validation split and a Brier-score reward.
get_lora_param_count() now supports Inkling-Small.
Changes
None.
Fixed
Accepts model names with a variant suffix, such as :peft:262144, in get_lr(), get_lora_param_count(), and get_full_finetune_param_count(). Before, these names raised an error.
Use Qwen/Qwen3.8-27B with the qwen3_8_xhigh_reasoning, qwen3_8_medium_reasoning, qwen3_8_low_reasoning, or qwen3_8_disable_thinking renderers through get_renderer().
Changes
None.
Fixed
Groups consecutive tool responses into one user turn in the Qwen3.5, Qwen3.6, and Nemotron renderers. Qwen3.5 and Qwen3.6 also omit the blank line before <tool_call> when an assistant message has no text. This changes rendered tokens for conversations with tool calls.
Treats Qwen3 turn markers in user-provided message text as content rather than conversation boundaries. This changes rendered tokens for such text.
Deprecations
None.
Breaking changes
When using a reasoning-disabled Qwen, Kimi K2.5 and K2.6, or Nemotron renderer, remove reasoning from affected supervised targets or select a thinking-enabled renderer. Inputs previously accepted with a mismatched prompt can now raise RendererError.
Use nemotron3_preserve_thinking or nemotron3_ultra_preserve_thinking with get_renderer() to keep historical reasoning in multi-turn contexts.
Use nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 with the nemotron3_ultra, nemotron3_ultra_disable_thinking, and nemotron3_ultra_preserve_thinking renderers. get_lr() has no calibrated learning rate for it, so set one explicitly.
Call Renderer.build_supervised_examples() on renderers without the sequence-extension property to train on every assistant turn. Before, these renderers raised NotImplementedError.
Changes
Inkling rendering is now part of the base installation, and the [inkling] extra is removed. Installing tinker-cookbook[inkling] still works but prints a warning, so remove the extra from install commands and requirement files.
Fixed
Corrects the final Qwen3 assistant turn: reasoning gets its own lines inside the thinking block, a reply without reasoning gets an empty block, and a tool-only reply has no newline before <tool_call>. This changes rendered tokens and parsed reasoning.
Trims leading and trailing whitespace from string content in the Qwen3.5, Qwen3.6, and Nemotron renderers. This changes rendered tokens for such content.
Preserves content whitespace when the role_colon renderer parses its rendered output.
Excludes prompt-supplied reasoning-prefill tokens from supervised loss. Qwen3.5 also preserves the prompt boundary when adjacent newlines would otherwise merge into one token. This changes training weights for the reasoning modes of the Qwen3.5, Nemotron, DeepSeek, and Kimi K2.5 and K2.6 renderers.
Includes sampling-compatible thinking delimiters in DeepSeek thinking-mode supervised examples without reasoning.
Preserves the required opening thinking token when a custom prefill is supplied to the Kimi K2.5 and K2.6 renderers, placing the prefill after it.
Deprecations
None.
Breaking changes
Qwen3 with thinking disabled now rejects a nonempty supervised assistant target when the conversation contains no user message. Add a user message or use a thinking-enabled renderer instead of relying on the previous mismatched training prompt.
Use tinker_cookbook.capture to record training and sampling tokens, log probabilities, and run coordinates. Call capture_to_store(...) for a local SQLite store with HTTP queries and live SSE updates, or pass your own capture sink. For agent harnesses that do not call the SDK, run the Anthropic- and OpenAI-compatible proxy with python -m tinker_cookbook.capture.proxy.serve.
Use thinkingmachines/Inkling-Small with the model registry and the tml_v0 renderer. It gets the same agentic rollout defaults as Inkling, and get_tokenizer() returns the Inkling tokenizer for it.
Changes
The inkling.sample_reasoning script now maps minimal, low, and medium to 0.1, 0.2, and 0.7. Its smallest default generation budget applies below effort 0.2.
Fixed
Uses a default max_tokens of 2048 for Inspect evaluations, avoiding response truncation at 128 tokens when no limit is specified. This changes scores for such evaluations.
Keeps the Search-R1 worked example within its one-to-five-word answer instruction. This changes the recipe's prompt tokens.
Deprecations
None.
Breaking changes
set_rollout_executor() and the rollout_executor argument of rl.train now raise ValueError for a process pool that uses fork, because such workers could hang after inheriting a Tinker client. On Linux before Python 3.14, fork is the default, so create the pool with ProcessPoolExecutor(mp_context=multiprocessing.get_context("spawn")). The cookbook's own sweep pools now use spawn, so put parallel sweep.run() calls inside if __name__ == "__main__":. An unguarded call now raises RuntimeError.
The inkling.sample_reasoning script now chooses a generation budget of 4096, 8192, or 16384 tokens according to reasoning effort. Set max_tokens to use one budget for every effort; the script also reports when a response reaches the limit.
Use Thinking Machines’ Inkling model with the tml_v0 renderer. Pass effort from 0 up to but excluding 1 to build_generation_prompt() or build_supervised_examples() to control reasoning effort.
Use the recipes.audio speech transcription and emotion recognition recipes for supervised and reinforcement learning. The scripts.inkling sampling examples cover reasoning, images, and audio.
Pass a RolloutConfig to rl.rollout_runner.run_rollout() to set turn, token, tool-call, and time limits and configure parse-error retries. Start with the agentic() preset for multi-turn tool use.
Pass a tokenizer to NLLEvaluator to report bits per byte alongside loss. Supervised training enables this metric for its test dataset and excludes special tokens from the byte-normalized score.
Use inference.modal.prepare and inference.modal.serve to deploy a trained checkpoint as a persistent OpenAI-compatible SGLang endpoint on Modal.
Changes
None.
Fixed
Treats turns containing only malformed tool calls as parse failures in AgentToolMessageEnv, applying failed_parse_reward instead of grading them as completed answers. This changes rewards for those rollouts.
Deprecations
None.
Breaking changes
None.
Dependencies
Adds the inkling extra with tinker>=0.23.0, tml-renderers>=0.0.1, and torch>=2.10.
Adds the audio extra with jiwer>=3.0.0 and soundfile>=0.12.0.
modal>=1.0.0 → >=1.5.1 in the modal, eval, eval-livecodebench, eval-mbpp, eval-swe-bench, eval-terminal-bench, and all extras.
Updates distillation recipe defaults to Qwen3.5 students and teachers, and Kimi K2.6 for multi-turn Harbor distillation. Override model_name and teacher_model to choose other supported models.
Fixed
Loads Kimi K2 tokenizers with their model-specific chat template, avoiding the Transformers fallback that breaks tool formatting.
Use Nemotron 3 Ultra with the nemotron3_ultra, nemotron3_ultra_medium_thinking, or nemotron3_ultra_disable_thinking renderer. The model registry also adds Qwen3.5-9B, Qwen3.5-9B-Base, and Qwen3.5-35B-A3B-Base.
Pass save_every_tokens or save_every_seconds to supervised.train.Config to schedule checkpoints by processed tokens or elapsed time. Custom loops using CheckpointManager can pass elapsed_tokens to maybe_save_async().
Use python -m tinker_cookbook.scripts.save_audit_log to archive audit logs as daily JSONL files. Pass --once for a single fetch or --backfill-days to include earlier days; your account must have permission to read audit logs.
Pass --destination-project-id to scripts.copy_checkpoint to choose the project receiving a copied checkpoint. Copies within your organization can use your current credentials without a separate --source-access-token.
Changes
Updates recipe default models from Llama and Qwen models to Qwen3.5, and from Kimi K2 Thinking to Kimi K2.6. Pass model_name to choose a different supported model.
Fixed
Preserves tool calls and tool results in the Inspect AI integration and honors required or named tool_choice settings.
Deprecations
None.
Breaking changes
None.
Dependencies
tinker>=0.9.0 → >=0.22.3.
Adds the tutorials extra with marimo>=0.23.8, matplotlib>=3.7.0, math-verify>=0.5.0, pylatexenc>=2.0, and sympy>=1.12.0.
Loads the Kimi K2.6 tokenizer with its required vocabulary files and model-specific tool-call chat template.
Deprecations
None.
Breaking changes
get_lora_param_count() now returns a single parameter count that accounts for the selected training components. Replace detailed, include_experts, and shared_expert_outer_loras with train_mlp, train_attn, and train_unembed flags matching your training client; dictionary results are no longer available.
Use Kimi K2.6 and Qwen3.6 models in Cookbook recipes. Kimi K2.6 includes thinking, non-thinking, and reasoning-history-preserving renderers; Nemotron 3 Super adds the nemotron3_low_thinking renderer.
Pass submit_ahead to supervised.train.Config to control how many training steps are submitted ahead, and async_periodic_saves=True to save periodic checkpoints in the background.
Use reverse=True with topk>0 in the SDFT recipe to train with analytical reverse KL over the teacher’s top-K tokens.
Use python -m tinker_cookbook.scripts.copy_checkpoint to copy a training checkpoint into your account. Pass --source-path and --source-access-token; set --output-kind training to keep training from the copy.
Use recipes.multiplayer_rl.text_arena.play to play against a trained tic-tac-toe model or evaluate it against random and optimal opponents.
Changes
conversation_to_datum() now normalizes loss weights per example after truncation, so examples with more trainable tokens no longer receive more total weight. Pass reduction="none" to retain unnormalized token weights.
Fixed
Handles sandbox creation failures during Harbor evaluation and uploads nested fixture directories without flattening their paths.
Deprecations
None.
Breaking changes
Renderer.parse_response() returns a ParseTermination value instead of a success boolean. Unpack message, termination = renderer.parse_response(tokens) and check termination.is_clean, or termination.is_stop_sequence for strict stop-sequence validation. All enum values are truthy, including MALFORMED; a best-effort message is still returned for malformed responses.
Replace RollingCheckpointManager with CheckpointManager and pass scheduling options by keyword. The manager handles periodic and rolling saves; use save_final_async(loop_state) to write a permanent final checkpoint and clean up rolling checkpoints.