Ember-1
Ember-1 is Fireworks Research's first custom post-trained model: a Kimi K3 derivative re-trained to emit roughly 40% fewer reasoning tokens at comparable quality, shipping with a 1,048,576-token context window, text+image input, and $3/$15 per-million-token pricing. It is a research preview, not a new architecture — and its only published third-party scores are the Vals AI real-world agentic suites (Vibe Code Bench v1.1 83.9%, CyberBench v1.1 69.5%, Finance Agent v2 51.9%, Terminal-Bench 4.0 ~20.7%).
Empirical Evaluation Results
Architectural Profile & Capabilities
IDENTITY & PROVENANCE Ember-1 (OpenRouter slug fireworks/ember-1) was announced by Fireworks Research on 23 September 2026 as the first entry in a custom model series. It is explicitly described by Fireworks as "built on Kimi K3" — Moonshot's frontier base model — and post-trained to generate shorter reasoning traces. Fireworks' headline claim is "Ember-1: half the tokens, same answers," operationalized as roughly 40% fewer tokens across Fireworks' own evaluations while maintaining comparable quality. Public material establishes the base model and the token-efficiency objective but does not disclose the full recipe (teacher checkpoints, loss function, preference data, or whether compression was achieved via SFT on shortened traces, RL with a length penalty, or both). No Ember-1 parameter count, active-parameter count, or attention-mechanism spec has been published. ARCHITECTURE Because Ember-1 is a Kimi K3 post-train rather than a new architecture, its compute graph is inherited: a sparse Mixture-of-Experts transformer with MLA-family attention and native long-context handling. Third-party reporting puts the underlying Kimi K3 at roughly 2.78 trillion total parameters in a sparse MoE configuration; that figure belongs to the base model and is NOT an independently published Ember-1 specification, so treat active-parameter estimates as inherited inference, not vendor-confirmed fact. Ember-1 is a reasoning model in the test-time-compute sense: it emits extended reasoning traces before answering, and its distinguishing property is that those traces are materially shorter than the K3 baseline rather than truncated at inference time. Multimodality is limited to image input (modality: image) alongside text; there is no audio or video path. CONTEXT & SERVING Context window is 1,048,576 tokens with a reported maximum output of 943,718 tokens — unusually large output headroom that makes single-call long-form synthesis and full-repository agentic loops feasible without chunk stitching. Serving is Fireworks serverless at $3.00/M input, $0.30/M cached input, $15.00/M output (priority tier $3.75/$0.375/$18.75). Third-party measurement puts throughput at approximately 51 tokens/second, consistent with a large MoE reasoning model rather than a latency-optimized small model — and because the billed denominator is output tokens, the 40% trace compression is the model's real economic thesis: fewer completion tokens at a flagship-tier output rate. MEASURED PERFORMANCE (EXACT VARIANT) The authoritative third-party evaluation is Vals AI, which added Ember-1 on release and ranks it #29 of 49 models on the Vals Index at 50.94%, with its best single result #17 of 113 on Vibe Code Bench v1.1. Vals' own summary is that across six benchmarks Ember-1 stays within roughly 3 points of Kimi K3 while using 11%–32% fewer reasoning tokens per task (note: Vals' independent measurement of token savings is far more conservative than Fireworks' ~40% claim). The profile is lopsided rather than uniformly strong: excellent at end-to-end web application generation (Vibe Code Bench v1.1 83.9%) and cyber tasks (CyberBench v1.1 69.5%), mid-pack on finance research (Finance Agent v2 51.9%), proof mathematics (ProofBench v1.1 61.0%) and classic algorithmic competition (IOI 46.3%), and weak on terminal-native agentic work (Terminal-Bench 4.0 ~20.7%) and truly catastrophic on Harvey's Legal Agent Benchmark (9.2%) — the latter indicating that aggressive trace compression buys cost efficiency at the expense of deep, multi-document professional reasoning. COMPETITIVE POSITION (OCTOBER 2026) Ember-1 is not a frontier-intelligence play. It should be read as a cost-efficiency derivative aimed at the GPT-6 Astra / Claude Opus 5 / Gemini 3.8 Pro / Kimi K3 tier of *workflows* while sitting a full band below them on hard reasoning. Against the flash/mini tier (Gemini 3.8 Flash, DeepSeek V4.1 Flash, Qwen 3.8) it competes on long-context depth and image input rather than on price — $15/M output is 20x+ DeepSeek V4.1 Flash's $0.66/M, so the economics only clear when the 40% trace reduction and 1M context actually eliminate retries and chunking. Fetching the same Kimi K3-derived capability from a cheaper Kimi K3 reseller can undercut it per-token, which is the primary procurement risk. Treat it as a research preview: benchmark coverage is one evaluator's suite, no vendor model card with architectural tables exists, and no Artificial Analysis or SWE-bench Verified figures have been published for this exact variant.
Ember-1 is not a new architecture — Fireworks Research describes it as a Kimi K3-based post-train optimized for reasoning-trace compression (~40% fewer tokens per Fireworks; 11%–32% measured independently by Vals AI). It therefore inherits Kimi K3's compute graph: a sparse MoE transformer with MLA-family attention and native long-context extrapolation. Third-party reporting attributes ~2.78T total parameters to the underlying Kimi K3; no Ember-1-specific total or active-parameter count has been published by Fireworks, so active-parameter figures should be treated as inherited inference rather than vendor specification. Context window 1,048,576 tokens with reported max output of 943,718 tokens. Input modality is text + image (no audio/video). Test-time reasoning is on by default and generates a visible/served reasoning trace before the final answer; the model's differentiator is trace length rather than trace presence. Served serverless on Fireworks at $3.00/M input, $0.30/M cached input, $15.00/M output; measured throughput ≈51 tok/s.
Recommended Workloads & Primary Use Cases
- •Published benchmark coverage is thin and single-sourced: essentially all exact-variant numbers come from Vals AI's real-world agentic suites. There are no published scores for Ember-1 on GPQA Diamond, HLE (with tools) or HLE (no tools), AIME 2026, SWE-bench Verified or SWE-bench Pro, ARC-AGI-2, BrowseComp, TAU-2/TAU-3, MRCR v2, or AA Intelligence Index v3.0/v4. Widely circulated figures such as SWE-bench Verified 92.2%, Terminal-Bench 2.1 82.0%, and DeepSWE 75.2% appear only in unverified secondary social posts framed as comparisons against 'K3 Max' and could not be attributed to Ember-1 — do not plan against them.
- •The headline 40% token-efficiency claim does not replicate at that magnitude independently: Vals AI measured only 11%–32% fewer reasoning tokens per task across six benchmarks while staying within roughly 3 points of Kimi K3. Budget cost savings at the conservative end, and validate trace length on your own prompts before committing to a unit-economics model.
- •Weak professional-agentic depth: Harvey's Legal Agent Benchmark at 9.2% and Terminal-Bench 4.0 at approximately 20.7% mean the model should not be trusted with multi-document legal reasoning or terminal/CLI-driven autonomous agents without human checkpoints. Note also a reporting discrepancy on Terminal-Bench 4.0 (20.7% on the OpenRouter/Vals listing versus 19.7% on BenchLeader) — treat it as low-20s, not a precise figure.
- •Latency and time-to-first-token are hostile to interactive UX: measured throughput is roughly 51 tok/s and Vals recorded 22m02s of wall-clock per task on CyberBench. With reasoning traces on by default on a ~2.8T-parameter MoE, TTFT for long-context prompts (hundreds of thousands of tokens) will be substantial. This is a batch/background model, not a chat model — decouple it from synchronous request paths and stream or poll.
- •Commercial and lifecycle risk: Ember-1 shipped as a research preview on 23 September 2026 with no published architecture card, no parameter disclosure, and no fine-tuning or weight access. Pricing ($3/$15, cached $0.30) is flagship-tier output pricing for a model that trails frontier models on hard reasoning, and the identical Kimi K3 capability is available from cheaper resellers — so evaluate switching cost, deprecation notice policy, and prompt-format/tool-calling stability (inherited Kimi tool-call schema) before hard-coding it into production.
Calls to Ember-1 are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by Fireworks.